US20100094970A1 - Latency based selection of fractional-storage servers - Google Patents

Abstract

Latency based selection of fractional-storage servers, including the steps of identifying a first group of fractional-storage servers estimated to have low response latencies in relation to an assembling device. Retrieving, by the assembling device from a second group of fractional-storage servers, enough erasure-coded fragments for reconstructing approximately sequential segments of streaming content. While retrieving the fragments, identifying at least one server from the second group having latency higher than a certain threshold in response to a fragment pull protocol request. And using the fragment pull protocol to replace the identified server with at least one server selected from the first group.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS
• This application claims the benefit of U.S. Provisional Patent Application No. 61/105,683, filed Oct. 15, 2008, and U.S. Provisional Patent Application No. 61,251,437, filed Oct. 14, 2009.
BACKGROUND
• A wide-area distributed storage system typically includes many servers at different locations. Due to its nature, a client communicating with a plurality of servers may encounter different latency values to the different servers, different load levels, and other parameters characterizing servers belonging to the distributed system, which may have an adverse effect on system responsiveness and latency.
BRIEF SUMMARY
• In one embodiment, a method for retrieving erasure-coded fragments, comprising: identifying a first group of fractional-storage servers estimated to have low response latencies in relation to an assembling device; retrieving, by the assembling device from a second group of fractional-storage servers, enough erasure-coded fragments needed for reconstructing approximately sequential segments of streaming content; while retrieving the fragments, identifying at least one server from the second group having latency higher than a certain threshold in response to a fragment pull protocol request; and using the fragment pull protocol to replace the identified server with at least one server selected from the first group.
• In one embodiment, a system comprising: a plurality of fractional-storage CDN servers connected to the public Internet; the servers store, at an average storage gain>5, erasure-coded fragments associated with approximately sequential segments of streaming contents, and are configured to respond with fragments to fragment pull protocol requests issued by assembling devices; wherein the system is configured to direct the fragment pull protocol requests of each assembling device towards servers exhibiting relatively low latencies in responding to the assembling device.
• In one embodiment, a system comprising: a plurality of fractional-storage CDN servers accessed via the public Internet and spread over a wide area; the servers store, at an average storage gain>5, erasure-coded fragments associated with approximately sequential segments of streaming contents, and are configured to deliver sub-transmissions comprising fragment sequences to assembling devices; wherein there is a high variance in latency between the servers and the assembling devices due to the fact that the system is spread over a wide area, and the system is configured to match assembling devices with servers having relatively low response latencies for sub-transmissions delivery.
• Implementations of the disclosed embodiments involve performing or completing selected tasks or steps manually, semi-automatically, fully automatically, and/or a combination thereof. Moreover, depending upon actual instrumentation and/or equipment used for implementing the disclosed embodiments, several embodiments could be achieved by hardware, by software, by firmware, or a combination thereof. In particular, with hardware, embodiments of the invention could exist by variations in the physical structure. Additionally, or alternatively, with software, selected functions of the invention could be performed by a data processor, such as a computing platform, executing software instructions or protocols using any suitable computer operating system. Moreover, features of the embodiments may be combined.
BRIEF DESCRIPTION OF THE DRAWINGS
• The embodiments are herein described, by way of example only, with reference to the accompanying drawings. No attempt is made to show structural details of the embodiments in more detail than is necessary for a fundamental understanding of the embodiments. In the drawings:
• FIG. 1 illustrates retrieving fragments according to locality.
• FIG. 2 is a flow diagram of one method in accordance with one embodiment.
• FIG. 3 illustrates one embodiment of segmenting content, encoding the segments into erasure-coded fragments, distributing the fragments to fractional-storage servers, and obtaining the fragments by assembling devices and assembling servers.
• FIG. 4 illustrates distribution and storage of erasure-coded fragments on fractional-storage servers.
• FIG. 5 illustrates three examples of changes made to redundancy factors according to changes in demand.
• FIG. 6 illustrates an assembling device obtaining erasure-coded fragments from fractional-storage servers.
• FIG. 7 illustrates real time fragment retrieval, segment reconstruction, and content presentation.
• FIG. 8 illustrates fractional-storage servers having the same bandwidth capability.
• FIG. 9 illustrates fractional-storage servers having different bandwidth capabilities.
• FIG. 10 andFIG. 11 illustrate a case where a fractional-storage server has failed.
• FIG. 12 illustrates a server failure due to network congestion.
• FIG. 13 toFIG. 15 illustrate changes in content consumption.
• FIG. 16 illustrates utilization of the entire aggregated bandwidth of fractional-storage servers for multiple content delivery.
• FIG. 17 toFIG. 20 illustrate various embodiments of fragment pull protocols.
• FIG. 21 andFIG. 22 illustrate different embodiments of content segmentation.
• FIG. 23 illustrates geographically distributed fractional-storage servers.
• FIG. 24 illustrates peak-to-average traffic ratios generated by assembling devices distributed over different time zones.
• FIG. 25 illustrates US-based fractional-storage servers delivering erasure-coded fragments to assembling devices spread over the globe.
• FIG. 26 illustrates different loads at different times for different time zones.
• FIG. 27 illustrates data centers communicating via shared links.
• FIG. 28 illustrates fractional-storage servers communicating via shared networks.
• FIG. 29 toFIG. 31 illustrate the influence of selecting source servers on backbone traffic.
• FIG. 32 illustrates server selection for network path determination.
• FIG. 33 illustrates boosting fractional-storage servers' bandwidth using P2P devices.
• FIG. 34 illustrates fractional-storage servers located on the Internet backbone.
• FIG. 35 illustrates an assembling server located at a network juncture.
• FIG. 36 illustrates fractional-storage servers placed at different locations.
• FIG. 37 illustrates one embodiment where a data center hosting fractional-storage servers has failed and is replaced by a different data center.
• FIG. 38 andFIG. 39 illustrate operation of multi data-center CDN.
DETAILED DESCRIPTION
• FIG. 1 illustrates one embodiment of a fractional-storage system storing erasure-coded fragments. Content, which may optionally be streaming content, is segmented into content segments. Each of the segments is encoded into erasure-coded fragments. The erasure-coded fragments are distributed to groups of fractional-storage servers242,248 or249. The erasure-coded fragments are then obtained by assemblingdevices235,237, and238. The obtained erasure-coded fragments are decoded to reconstruct the segments.
• In one embodiment, a method for selecting enough new servers from which to obtain fragments, based on latency, includes the following steps: (i) accessing data regarding the relevant servers; (ii) accessing data regarding the latencies from the relevant servers to the assembling device; and (iii) obtaining fragments from enough of the relevant servers having the lowest latencies; or obtaining fragments from enough of the relevant servers selected randomly and having latencies below a certain threshold.
• In one embodiment, a method for selecting enough new servers from which to obtain fragments, based on bandwidth and latency, includes the following steps: (i) accessing data regarding the relevant servers; (ii) accessing data regarding the unutilized bandwidth of the relevant servers; (iii) identifying more than enough relevant servers having the most unutilized bandwidth; or randomly identifying more than enough relevant servers having unutilized bandwidth above a certain threshold; (iv) accessing data regarding the latencies from the identified servers to the assembling device; and (v) obtaining fragments from enough of the identified servers having the lowest latencies; or obtaining fragments from enough of the relevant servers selected randomly and having latencies below a certain threshold.
• In one embodiment, a method for selecting enough new servers from which to obtain fragments, based on latency and bandwidth, includes the following steps: (i) accessing data regarding the relevant servers; (ii) identifying more than enough relevant servers having latencies to the assembling device below a certain threshold; or randomly identifying more than enough relevant servers having latencies to the assembling device below a certain threshold; (iii) accessing data regarding the unutilized bandwidth of the identified servers; and (iv) obtaining fragments from enough of the identified servers having the highest unutilized bandwidth; or obtaining fragments from enough of the relevant servers selected randomly and having the highest unutilized bandwidth.
• In one embodiment, a method for selecting enough new servers from which to obtain fragments, based on latency and locality, includes the following steps: (i) accessing data regarding the relevant servers; (ii) identifying more than enough relevant servers having latencies to the assembling device below a certain threshold; or randomly identifying more than enough relevant servers having latencies to the assembling device below a certain threshold; (iii) accessing data regarding the network topology distance from the relevant servers to the assembling device; and (iv) obtaining fragments from enough of the topologically closest relevant servers; or obtaining fragments from enough of the relevant servers that are located in the same sub-network as the assembling device, or located in the closest sub-networks.
• In one embodiment, a method for selecting enough new servers from which to obtain fragments is based on bandwidth, latency, locality, and, optionally, one or more additional relevant parameters. The method may weigh the different parameters in various ways, all of them are intended to be covered by the embodiments. For example, the method may include the following steps: (i) accessing data regarding the relevant servers; (ii) receiving data regarding the unutilized bandwidth latencies to the assembling device, and topology distances to the assembling device; (iii) weighting the received data and identifying a quantity of the most proper relevant servers, which can provide enough fragments to reconstruct content; and (iv) obtaining the fragments from the identified servers. In another example, the method may include the following steps: (i) accessing data regarding the relevant servers; (ii) identifying a set of more than enough relevant servers having the most unutilized bandwidth; or randomly identifying a set of more than enough relevant servers having unutilized bandwidth above a certain threshold; (iii) from the set, identifying a sub-set of more than enough relevant servers having latencies to the assembling device below a certain threshold; or randomly identifying more than enough relevant servers having latencies to the assembling device below a certain threshold; and (iv) obtaining fragments from enough of the topologically closest relevant servers out of the sub-set; or obtaining fragments from enough of the relevant servers out of the sub-sets, which are located in the same sub-network as the assembling device, or located in the closest sub-networks.
• In one embodiment, a server may be loaded to a point that it is approximately unable to transmit additional fragments as a response to new fragment requests or new sub-transmission requests. The server may also be too loaded to continue transmitting fragments to its currently served assembling devices. In one example, these cases result from one or more of the following conditions: (i) server hardware limitation, such as CPU power or memory bus constraints, which prevents it from delivering fragments beyond a certain throughput, (ii) outgoing communication link limitation, such as a fixed-bandwidth line, which prevents the server from transmitting fragments beyond a rate that can be supported by the line, (iii) sharing of an outgoing communication line with other servers, and the other servers utilizing the shared line to a point that lowers the bandwidth available for fragment transmission, and (iv) sharing the fragment storage and transmission software together with other applications on one physical server, and the other applications consuming CPU, memory, or communication resources to a point that affects the ability of the fragment storage and transmission software to respond to fragment or sub-transmission requests.
• FIG. 2 is a flow diagram illustrating one method for retrieving erasure-coded fragments, comprising the following steps: Instep7240, identifying a first group of fractional-storage servers estimated to have low response latencies in relation to an assembling device. In step7241, retrieving, by the assembling device from a second group of fractional-storage servers, enough erasure-coded fragments needed for reconstructing approximately sequential segments of streaming content. Optionally, measuring the round trip delay between the assembling device and the servers using the pull protocol requests. Optionally, identifying the servers of the first group is performed by obtaining the router hop-count between the servers and the assembling device, whereby a lower hop-count usually indicates lower latency. Instep7242, while retrieving the fragments, identifying at least one server from the second group having latency higher than a certain threshold in response to a fragment pull protocol request. And instep7243, using the fragment pull protocol to replace the identified server with at least one server selected from the first group. Optionally, the servers are fractional-storage CDN servers, and the at least one server selected from the first group is selected approximately randomly. Optionally, the servers are fractional-storage CDN servers, and the at least one server selected from the first group is selected according to a least latency criterion. Optionally, the servers are fractional-storage CDN servers, and the at least one server selected from the first group is approximately the least loaded server. Optionally, the method further comprising while retrieving the fragments, identifying at least one server from the second group having latency variance higher than a certain threshold, and using the fragment pull protocol to replace that server with at least one server selected from the first group. Optionally, identifying the at least one server having latency higher than a certain threshold is performed by the assembling device via the fragment pull protocol, whereby no dedicated communication is needed for identifying the server. Optionally, identifying the at least one server having latency higher than a certain threshold is performed by obtaining the router hop-count between the servers and the assembling device, whereby a higher hop-count usually indicates higher latency. Optionally, the erasure-coding is rateless-coding potentially resulting in fragments having a limitless redundancy factor; whereby a higher redundancy enables a larger first group. Optionally, identifying the servers of the first group is performed by occasionally measuring the latencies between the assembling device and the servers not belonging to the second group.
• The term “erasure coding” as used herein denotes a process in which a sequence of erasure-coded fragments can be generated from a segment such that the segment can be reconstructed from any or almost any subset of the erasure-coded fragments of size equal to or somewhat larger than the size of the segment (sometimes may be referred to as “enough erasure-coded fragments” or “sufficient subset of fragments”). Examples of erasure codes include, but are not limited to, rateless codes, Reed-Solomon codes, Tornado codes, Viterbi codes, Turbo codes, any Block codes, any Convolutional codes, and any other codes that are usually used for forward error correction (FEC).
• The term “rateless coding” as used herein denotes type of erasure coding in which a very long, potentially limitless, sequence of rateless-coded fragments can be generated from a segment such that the segment can be reconstructed from any or almost any subset of the rateless-coded fragments of size equal to or somewhat larger than the size of the segment (sometimes may be referred to as “enough rateless-coded fragments”). Examples of rateless codes include, but are not limited to, Raptor codes, LT codes, online codes, any Fountain codes, and any other Rateless codes.
• The term “erasure-coded fragment” denotes a fragment comprising data encoded with an erasure code (which may also be a rateless code in some embodiments). The term “rateless-coded fragment” denotes a fragment comprising data encoded with a rateless code.
• The term “assembling device” as used herein denotes a computing device that retrieves erasure-coded fragments from servers over a network. The assembling device may perform one or more of the following: (i) Decode the retrieved erasure-coded fragments into segments. (ii) Present the content reconstructed from the retrieved erasure-coded fragments. (iii) Act as a bandwidth amplification device, by receiving, storing, and forwarding erasure-coded fragments. In some embodiments, the assembling device may be any device located at the user premises, like an STB, PC, gaming console, DVD player, PVR device, or any other device able to retrieve erasure-coded fragments from a communication network. In some embodiments, the assembling device may be an assembling server. In some embodiments, the assembling device may be any computational device with access to a communication network, located at a central office, data center, BRAS location, ISP premises, or any other place with direct network connectivity. In one embodiment, the assembling device is coupled to a display device used for content presentation.
• The abbreviation CDN denotes “Content Delivery Network”. The term “CDN server” as used herein denotes a server having one or more of the following characteristics: (i) A bandwidth (CDN_BW) that is much greater than the average bandwidth consumed by a user premises device (User_BW) receiving video streaming content. In some examples, the CDN_BW is at least 10 times, 100 times, 1,000 times, or 10,000 times greater than the User_BW. (ii) The server is located outside the last mile communication infrastructure of the end users, such that the CDN server and the end users are located in different networks. For example, the CDN server is not located under a BRAS, while the end users are located under a BRAS. Moreover, in some embodiments, the CDN servers are deployed over a wide area across the Internet and optionally located close to or on the Internet backbone. In some embodiments, the CDN server does not usually retrieve and play streaming content. In some embodiments, the CDN server has a much greater storage space than the storage space of an average player of streaming content.
• The term “fractional-storage server” in the context of erasure-coded fragments (also applicable to “fractional-storage CDN server”), as used herein denotes a server that (i) stores less than the minimum quantity of erasure-coded fragments required to decode the erasure-coded fragments, and (ii) where at least a meaningful quantity of the stored erasure-coded fragments is not stored in order to be consumed by the fractional-storage server.
• The term “streaming content” as used herein denotes any type of content that can begin playing as it is being delivered. Streaming content may be delivered using a streaming protocol, a progressive download protocol, or any other protocol enabling a client to begin playing the content as it is being delivered. Moreover, the term “streaming protocol” includes “progressive download protocol”. In addition, the verb “streaming” refers to using a streaming protocol, using a progressive download protocol, or using any other protocol enabling the receiver to begin playing the content as it is being delivered.
• In one embodiment, assembling devices measure the latency of the different servers in responding to fragment requests, and then use the latency information to estimate the loads on the servers. In one example, a high latency may indicate a high load on the server.
• In one embodiment, the topological router hop count between an assembling device and fragment delivering servers is used to estimate the latency of the servers in responding to fragment requests.
• In one embodiment, the latency of fragment delivering servers in responding to fragment requests by an assembling device is used to estimate the topological router hop count between an assembling device and the servers.
• In one embodiment, the assembling devices perform several latency measurements for the different servers in responding to fragment requests, and then use the latency variance information to estimate the loads on the servers. In one example, a high latency variance may suggest a high load on server.
• FIG. 1 further illustrates a few examples of retrieving fragments according to locality. In one example, the fractional-storage servers are connected to a data network or networks comprising therouters201 to209. Assemblingdevices235,237, and238 are connected to the same data network or networks, and K=3, meaning that any assembling device needs to obtain 3 erasure-coded fragments per segment from optionally 3 different servers out of the 10 in order to successfully reconstruct the segment.
• Each assembling device tries to obtain erasure-coded fragments from fractional-storage servers that are closest to it topologically. In one embodiment, the topological distance is a function of the number of separating routers. Assemblingdevice238 can select three servers fromgroups242,248 or249. According to the minimal path criterion, it retrieves the erasure-coded fragments fromservers399hto399iofgroup248, since they are only onerouter208 away.Groups242 and249 are three (208,202,203) and five (208,202,203,201,209) routers away, and are therefore not selected for retrieval. Similarly,device237 selects three servers out ofgroup242, anddevice235 can select any three servers fromgroups242 and249, since both are located four routers away.
• In one embodiment, if topologically close servers do not respond to the assembling device, or report a bandwidth limitation, the assembling device will attempt to obtain an erasure-coded fragment from the next topologically closest server.
• In one embodiment, an assembling device attempts to obtain erasure-coded fragments from servers featuring the lowest latency. Upon no response, for whatever reason, the assembling device will attempt to retrieve from the next lowest latency server. In one embodiment, the assembling device obtains information regarding the unutilized fragment delivery bandwidths of servers, and then attempts to retrieve from the lowest latency servers out of the servers having enough unutilized bandwidth. In one embodiment, the assembling device obtains information regarding the unutilized fragment delivery bandwidths of the servers, and then attempts to retrieve from the topologically closest servers out of the servers having enough unutilized bandwidth.
• Still referring toFIG. 1, in one embodiment the assembling devices select servers according to a latency criterion, such as selecting servers with the shortest time between fragment request and fragment delivery, or selecting servers having latency below a dynamic or static threshold. Assemblingdevice237 assembles content fromservers399c,399f,399g, and assemblingdevice235 assembles content fromservers399b,399c,399g(both use a mixture of servers fromgroups242 and249). At a certain point in time,router209 becomes congested or blocked, and prevents the erasure-coded fragments fromservers399band399cfrom arriving at assemblingdevices235 and237, or causes the fragments to arrive with an increased delay. Therefore, assemblingdevice235 switches to three servers ofgroup242, and assemblingdevice237 switches fromserver399ctoserver399e.
• In one embodiment, the assembling device selects fractional-storage servers according to the following criterion: first, servers with adequate unutilized fragment delivery bandwidth are considered, then out of these, those with latency below a threshold are considered, and out of these, the servers with minimal topological routing path are selected.
• In some embodiments, the assembling devices use a fragment pull protocol to retrieve the fragments, and approach servers having low latency or low hop count as compared to other servers. In some embodiments, the assembling devices use a push protocol to retrieve the fragments, and approach servers having low latency or low hop count as compared to other servers, optionally by obtaining multiple sub-transmissions comprising fragment sequences.
• FIG. 3 illustrates one embodiment of a fractional-storage system storing erasure-coded fragments.Content100, which may optionally be streaming content, is segmented intocontent segments101a,101bto101j(for brevity referred to as segments). Each of the segments is encoded into erasure-coded fragments. For example,segment101ais encoded into erasure-codedfragments390ato390(N). The erasure-coded fragments are distributed to the fractional-storage servers399ato399(N) and/or to the bandwidth amplification devices610aa. The erasure-coded fragments are then obtained by assembling devices like661 or proxy servers likeproxy server661sfrom the fractional-storage servers399ato399(N) and/or the bandwidth amplification devices610aa. The obtained erasure-coded fragments are decoded to reconstruct the segments. Theproxy server661smay broadcast/multicast and/or re-stream the reconstructed content, optionally using standard streaming technique, to its client(s)661o, optionally overnetwork300n. In some embodiments, the content distribution is performed in real time. In some embodiments, the content assembly is performed in real time and the presentation starts a short time after the content request.
• Similarly tocontent100, additional contents are segmented, encoded into erasure-coded fragments, and distributed to the fractional-storage servers and/or to the bandwidth amplification devices. Each segment may be reconstructed independently of other segments by obtaining and decoding enough erasure-coded fragments generated from that segment.
• In some embodiments, the encoding scheme is erasure codes and/or rateless codes. In some embodiments, the fractional-storage servers399ato399(N) are Content Delivery Network (CDN) servers, optionally accessed over the public Internet. In some embodiments, the control, management, content reception, content segmentation, segment encoding, erasure-coded fragment distribution, allocation of bandwidth amplification devices, and/or other kind of central supervision and operation may be performed by managing server(s)393, which may be a part of the CDN network. It is noted that the term “fractional-storage server” is not limited to a large server and, according to the context, may include a fractional-storage bandwidth amplification device, a fractional-storage peer server, or other types of fractional-storage servers.
• FIG. 4 (without the fragments marked with dashed lines) illustrates one example of distributing the erasure-coded fragments to ‘M’CDN servers399ato399(M), connected to anetwork300. Encodedfragments310ato310(M) of a first segment are sent for storage inservers399ato399(M) respectively. Similarly, erasure-codedfragments320ato320(M) of a second segment are sent for storage inservers399ato399(M) respectively. In addition, other erasure-coded fragments associated with other segments of other contents, illustrated as erasure-codedfragments390ato390(M), are sent for storage inservers399ato399(M) respectively. The number of unique erasure-coded fragments from each segment that are stored on the servers (399ato399(M)) is equal to M in this example, where M may be smaller than the maximum number of unique erasure-coded fragments, meaning that only a subset of the potential erasure-coded fragments are actually stored. It is also possible to store the maximum number of unique erasure-coded fragments, or store more than one unique erasure-coded fragment per segment per server. Thenetwork300 may be the Internet for example, or any other data network connecting multiple nodes, such as a private IP network, or a Wide Area Network (“WAN”). In one embodiment, the fragments marked with dashed lines illustrate one example where (N-M) additional servers are added to the array, and (N-M) new unique erasure-coded fragments per segment per content (310(M+1) to310(N),320(M+1) to320(N), and390(M+1) to390(N)) are generated and added to the array. In one embodiment, only M out of the maximum possible erasure-coded fragments (L) are actually generated for storage in the first place. In one embodiment, when the additional N-M erasure-coded fragments are needed for storage (e.g., when additional servers are made available), the remainder of the N-M erasure-coded fragments are actually generated. Any time that additional unique erasure-coded fragments are needed, this process of calculating the additional erasure-coded fragments is repeated, up to the point that all L possible erasure-coded fragments are used.
• In one embodiment, and especially when using rateless coding, L may be chosen as a sufficiently large number to account for any realistic future growth of the server array. For example, a segment of 96 Kbytes is expanded using a rateless code with a ratio of 1 to 2̂16 original symbols to encoded data, into an encoding symbol of potential size 6.29 GBytes. Assuming a 1500 Bytes erasure-coded fragment size, then potentially 4.19 million unique erasure-coded fragments can be generated. Now, it is safe to assume that for all practical uses, the server array will not grow to more than 4.19 million nodes, and may contain several thousands of servers, meaning that the encoded data can be used in all cases where additional unique erasure-coded fragments are needed, by generating new erasure-coded fragments out of the segment. Optionally, a server may store erasure-coded fragments for only some of the segments.
• In one example of redundancy factor and storage gain (without the fragments marked with dashed lines),server399astores only erasure-codedfragment310afrom a first segment, erasure-codedfragment320afrom a second segment, and erasure-codedfragment390afrom a third segment. Assuming that: (i) the segment size is 1024 Kbytes; (ii) the segment is encoded using erasure code into a 4096 KByte encoded segment; (iii) the encoded segment is segmented into 256 erasure-coded fragments of size 4096/256=16 KByte; and (iv) the erasure-coded fragments are stored on 256 servers (M=256); it turns out that each server stores only a 1/64 portion of the original size of the segment. This means that each server can manage with only 1/64 of the storage requirements in comparison to a situation where it had to store the entire segment. In addition, there are 256 erasure-coded fragments altogether from each encoded segment, meaning that an assembling device that is assembling the erasure-coded fragments from the servers need only select slightly more than 64 erasure-coded fragments in order to completely reconstruct the segment, and it can select whichever slightly more than 64 erasure-coded fragments it desires out of the 256 possibly available. The redundancy factor in this example is approximately 256/64=4. All contents in this example enjoy a factor of 64 in storage gains, meaning thatserver399a, for example, stores only 1/64 of the information associated with the first segments and any additional segments belonging to other contents. In one example, each server supports high volume storage of between about 500 GByte and 500 TBytes, optionally utilizing hard drive, Solid State Drive, or any other high volume storage device(s). In these cases, each server may store many millions of erasure-coded fragments, associated with millions of segments, belonging to hundreds of thousands of different contents, and possibly more.
• In one embodiment, new content initially encoded with a low redundancy factor is distributed to an initial number of fractional-storage servers. As the content is distributed to more servers, additional unique fragments are encoded and therefore the redundancy factor increases. Optionally, as the content's popularity increases, and/or as the load on the fractional-storage servers increases, the redundancy factor is increased, and vice versa.
• In one embodiment, multiple unique erasure-coded fragments per segment of a new content are distributed to an initial number of fractional-storage servers with a low storage gain (i.e. each server stores multiple unique erasure-coded fragments per encoded segment). As the content is distributed to more fractional-storage servers, some of the erasure-coded fragments stored on the initial number of fractional-storage servers are removed and thereby the storage gain is increased. Optionally, as the demand for the content increases, the storage gain is decreased, and vice versa.
• The term “redundancy factor” as used herein denotes the following ratio: (total size of the unique erasure-coded fragments generated from a segment and actually stored on the servers)/(size of the segment).
• Assuming all segments have approximately the same size and all fragments generated from the segments have approximately the same size (without limiting any of the embodiments), the term “storage gain” as used herein denotes the following ratio: (size of a segment)/(size of an erasure-coded fragment). If the server stores more than one erasure-coded fragment per segment, the storage gain denotes the following ratio: (size of segment)/((size of erasure-coded fragment)*(number of stored erasure-coded fragments per segment)).
• FIG. 5 illustrates three examples (each depicted by one of the columns A-C) of changing the redundancy factor according to the demand. Column A illustrates one simplified example of a storage array including 16 servers (1001 to1016). Each server stores up to 2 different erasure-coded fragments, and can service an erasure-coded fragment transmission bandwidth of up to B. Assuming three contents (#1, #2, and #3) processed to segments and erasure-coded fragments with a storage gain of 4.
• Assumingcontent #1 is the most popular, and requires a peak bandwidth of 11×B. Since each server can service up to bandwidth B, at least 11 servers are needed to servicecontent #1 bandwidth requirements.Content #1 is therefore encoded into 11 unique erasure-coded fragments per segment, illustrated as group g1 of erasure-coded fragments stored onservers1001 to1011. Out of these 11 erasure-coded fragments, it is sufficient to obtain slightly more than 4 erasure-coded fragments in order to reconstruct a segment ofcontent #1. Therefore, the resulting redundancy factor of the stored fragments associated withcontent #1 is approximately 11/4=2.75.Content #2 requires less bandwidth, and manages with a peak of 7×B. It is therefore encoded into 7 unique erasure-coded fragments per segment, illustrated as group g2 of erasure-coded fragments onservers1010 to1016. Therefore, the redundancy factor of the stored fragments associated withcontent #2 is 7/4=1.75.Content #3 requires a peak bandwidth of 5×B, but for some reason (for example, being a more critical content), it is encoded into 14 erasure-coded fragments per segment, illustrated as group g3 of erasure-coded fragments onservers1001 to1009 and1012 to1016. Therefore, the redundancy factor of the stored fragments associated withcontent #3 is 14/4=3.5. This concludes the storage availability of the servers in this example, as every server stores two erasure-coded fragments.
• Column B illustrates an example wherecontent #2 becomes more popular thancontent #1, and therefore requires more bandwidth and hence more of a redundancy factor. This is achieved by eliminating 5 erasure-coded fragments associated withcontent #1 that were previously stored onservers1001 to1005, and replacing them with 5 new unique erasure-coded fragments g4 associated withcontent #2. This brings the total number of erasure-coded fragments per segments ofcontent #1 and #2 to 6 and 12 respectively. In column C,new content #4 is stored onservers1001 to1003 and1014 to1016 (illustrated as g5), by eliminating 3 erasure-coded fragments ofcontent #1 and 3 erasure-coded fragments ofcontent #2.
• Throughout the examples ofFIG. 5, a record of “what erasure-coded fragments are stored where” may be: (i) kept in each of theservers1001 to1016. In this case, when an assembling device is assemblingcontent #2, it will send a query toservers1001 to1016, asking which one is storing erasure-coded fragments ofcontent #2; (ii) kept in a control server. In this case, an assembling device will ask the control server to send back a list of all servers storing erasure-coded fragments of its required content.
• FIG. 6 illustrates one embodiment of a serverarray including servers399ato399(N) storing erasure-codedfragments390ato390(N) associated with content. In order for assemblingdevice661 to reconstruct asegment101aof the content, it has to retrieve at least K erasure-coded fragments. In one example, k=4 and theassembling device661 chooses approximately randomly from which servers to retrieve the 4 different erasure-coded fragments. It chooses to retrievefragments390a,390c,390(N−1) and390(N), which are noted asgroup573, and reconstruct thesegment101a. Consequent segments of the content are reconstructed in a similar fashion, and the content may eventually be fully retrieved by combining all relevant segments. If theassembling device661 cannot reconstruct thesegment101a, it retrieves one or more additional unique erasure-coded fragments, and tries again.
• The term “approximately random” as used herein refers to, but is not limited to, random, pseudo random, and/or based on a long list of numbers featuring very low autocorrelation and very low correlation with other similar lists of numbers.
• Referring back toFIG. 6, in one embodiment, the content being distributed supports stream presentation, andsegment101ais of small size, to enable content presentation by assemblingdevice661 shortly after beginning the reception of the segment (or any other segment of the content). For example,segment101ais 96 KByte, allowing a 5 Mbps download speed receiver to obtain the entire segment (by requesting enough erasure-coded fragments to enable the reconstruction of the segment, and such that the total size received of all requested erasure-coded fragments is slightly larger than the segment) after approximately 0.2 seconds from request, and beginning the presentation shortly or right after the successful decoding and reconstruction ofsegment101a.
• In some embodiments, the fragments are small enough to be contained in one packet. In one embodiment, each fragment is about 1400 bytes, and can fit into one UDP or RTP packet transmitted over Ethernet. The stateless nature of UDP and RTP allows the servers to send one packet with one fragment very quickly, without the need for any acknowledgement or hand shaking. In some embodiments, the fragment pull protocol requests use one stateless packet, like UDP or RTP. In one embodiment, the assembling device requests about 100 fragments approximately in parallel, using 100 separate requests or one or few aggregated requests. About 100 servers respond by sending about 100 fragments, each encapsulated in one stateless packet, after a short delay, and the assembling device receives the fragments within a fraction of a second. Assuming an Internet round trip delay of 100 ms, and server processing latency of 100 ms, then after 200 ms the assembling device starts receiving all 100 fragments. With a modem of 5 Mbps, and assuming 1400 bytes per fragment, all 100 fragments are received 1400×100×8/5 Mbps=224 ms after the initial delay, meaning that content can be presented200+224=424 ms after request (decoding and other process time has been ignored in this example).
• The following embodiments describe processes for on-the-fly erasure-coded fragment retrieval from fractional-storage servers.
• In one embodiment, a method for obtaining erasure-coded fragments from fractional-storage servers to reconstruct a segment includes the following steps: (i) identifying the next segment to be obtained; optionally, the segments are approximately sequential segments of streaming content obtained according to their sequential order; (ii) optionally, determining the minimum number of fragments needed to reconstruct the segment; (iii) are enough identified relevant servers (i.e. servers storing the required fragments) available from the process of obtaining prior segment/s? (iv) if no, identifying enough relevant servers; (v) if yes, requesting enough fragments from the identified relevant servers; if less than enough fragments are obtained from the identified relevant servers, go back to step iv and identify additional relevant server/s; (vi) reconstruct the segment from the obtained fragments; and (vii) optionally, go back to step i to obtain the next segment.
• In some embodiments, expressions like “approximately sequential segments” may denote one or more of the following non-limiting options: segments that are sequential (in time or according to a file's order), segments that are approximately sequential (such as segments with some interlace, or segments without a great amount of non-sequential data), segments generated sequentially and/or approximately sequentially from different components of content (such as storing the i-frames and p-frames of a compressed content in different segments), and/or other sequential or approximately sequential segmentation after classification or separation into different components and/or elements.
• In one embodiment, a method for obtaining erasure-coded fragments from fractional-storage servers to reconstruct multiple segments includes the following steps: (i) identifying multiple segments to be obtained, optionally according to their sequential order; (ii) optionally, determining the minimum number of fragments needed to reconstruct the segment; (iii) optionally, determining the number of fragments to be obtained approximately in parallel; (iv) are enough identified relevant servers available from the process of obtaining prior segment/s? (v) if no, identifying enough relevant servers; (vi) if yes, requesting enough fragments from the identified relevant servers, optionally in parallel and according to the sequential order of the segments; (vii) if less than enough fragments are obtained from the identified relevant servers, go back to step iv and identify additional relevant server/s; (viii) reconstructing the segment/s from the obtained fragments; and (ix) optionally, go back to step i to obtain the next segments.
• In one embodiment, a method for obtaining erasure-coded fragments from fractional-storage servers to reconstruct a segment in a burst mode includes the following steps: (i) identifying the next segment to be obtained; (ii) optionally, determining the minimum number of fragments needed to reconstruct the segment; (iii) are more than the minimum number of relevant servers available from the process of obtaining prior segment/s? (iv) if no, identifying more than the minimum relevant servers; (v) if yes, requesting more than the minimum number of fragments needed to reconstruct the segment; if less than enough fragments are obtained, go back to step iv and identify additional relevant server/s; (vi) reconstructing the segment from the obtained fragments; and (vii) optionally, go back to step i to obtain the next segment.
• The various methods for obtaining erasure-coded fragments from the fractional-storage servers for reconstructing one or more segments may be combined as needed. In one example, the initial segment/s are obtained using a burst mode and the following segments are retrieved without requesting extra fragments. In another example, the initial segment/s are obtained approximately in parallel and optionally using a burst mode, and the following segments are obtained one by one and optionally without requesting extra fragments. The fragments may be obtained using a pull protocol and/or a push protocol. Moreover, the servers from which to retrieve the fragments may be selected according to one or more of the various discussed methods for selecting the servers and/or load balancing the servers.
• FIG. 7 illustrates one embodiment of real time streaming content retrieval from fractional-storage servers. An assembling device begins a process of obtainingstreaming content700 for presentation. Starting at T1, the assembling device requests erasure-codedfragments720ato720(K). By T2, all K erasure-coded fragments are obtained, and at time T2buntil T4, erasure-codedfragments720ato720(K) are decoded intosegment710a. The retrieval time of the erasure-coded fragments and the segment decoding time should be equal to or faster than the corresponding presentation time, in order to enable a continuous presentation, once presentation begins at T5. T2bminus T2 is a short delay, and can be fractions of a second. Subsequent erasure-codedfragments730ato730(K) are retrieved between T2 and T3, and are decoded intosubsequent segment710bbetween T4 and T6.
• In one example, the streamingcontent700 is encoded at 1 Mbps, and the segment size is 96 Kbytes. The presentation of each segment takes about 0.77 seconds. Retrievingfragments720ato720(K) takes no more than 0.77 seconds, meaning that the assembling device's connection bandwidth must be 1 Mbps or higher. Decodingsegment710atakes no more than 0.77 seconds. If a small delay of 0.2 seconds is assumed for both T2bminus T2 and T5 minus T4, then T5 can start at 0.77+0.2+0.77+0.2=1.94 seconds after T1, meaning that presentation can begin about 2 seconds following request of the first erasure-coded fragment. In another example, the retrieval process and the decoding process are performed faster than the real time presentation bounds, therefore enabling a shorter time to play and a download rate that exceeds the presentation rate.
• In one embodiment, the erasure-codedfragments720ato720(K) are retrieved in approximately random order, or any other order, as long as at least the K erasure-coded fragments needed for decoding thesegment710aare available until time T2.
• In one embodiment, the fragments associated with sequential segments of streaming content are delivered to an assembling device as a plurality of sub-transmissions. In this case, each fractional-storage server participating in the delivery of the fragments to the assembling device sends a transmission to the assembling device comprising a sequence of erasure-coded fragments. This transmission is referred to as a sub-transmission. In one example, each sub-transmission contains at least one fragment per each sequential segment of the streaming content. In one example, the sub-transmission starts at a segment indicated by the assembling device, and continues from that point onwards, approximately according to the sequential order of segments, until the assembling device instructs the server to stop, or until reaching the last segment of the content. Each sub-transmission carries only a fraction of the fragments (per segment) needed to reconstruct the segments of the streaming content, such that the combination of at least two sub-transmissions received by the assembling device from the servers allows the assembling device to obtain enough fragments needed to reconstruct each segment.
• In one embodiment, each sub-transmission is delivered to the assembling device via a streaming session, such as an RTP session, wherein the RTP packets transport the fragment sequence approximately according to the order of the sequential segments. In one embodiment, each sub-transmission is delivered to the assembling device via an HTTP connection, or other closed-loop data transfer mechanisms over TCP/IP. In one embodiment, the assembling device may change one or more transmitting servers on the fly, by instructing the server(s) to stop sending an already active sub-transmission—as may be needed in a case of an RTP session, and initiating new sub-transmissions from other servers instead. Replacement of transmitting servers on the fly may be needed in a case of a server failure, network failure, or high load or latency conditions.
• In some embodiments, a push protocol is used to obtain fragments. A push protocol may be implemented using one transmission carrying fragments from a source server to a destination receiver, or may be implemented using a plurality of sub-transmissions. When using sub-transmissions, each sub-transmission transports a fraction of the fragments needed for segment reconstruction. Segments may be reconstructed from fragments received via sub-transmissions after obtaining decodable sets of erasure-coded fragments; optionally one set per segment. A sub-transmission may be transported using an IP stream such as RTP, an HTTPS session, or any other protocol suitable for transporting a sequence of fragments between a source server and a destination assembling device.
• FIG. 6 illustrates one embodiment, in which content is segmented and erasure-coded.Fragments390ato390(N), belonging to a first segment, are distributed toservers399ato399(N) respectively. Other fragments belonging to subsequent segments are similarly distributed toservers399ato399(N). The servers may use a push protocol to transport the fragments to an assembling device. A push protocol sub-transmission may comprise a sequence of fragments associated with multiple segments. In one example, the fragments are ordered according to the sequential order of the segments in a streaming content.Server399asends a first sub-transmission to a destination assembling-device. Optionally, the first sub-transmission comprises a sequence of fragments starting withfragment390a, associated with the first segment, and continuing with fragments belonging to subsequent segments.Server399csends a second sub-transmission to the destination assembling-device, optionally starting withfragment390c, associated with the first segment, and continuing with fragments belonging to subsequent segments. In a similar fashion, servers399(N−1) and399(N) send additional sub-transmissions to the destination assembling-device, each comprising a unique fragment sequence.
• When using a push transmission, the assembling device does not explicitly ask for each fragment, but instead instructs each of the different servers to start sending it a fragment sequence using a sub-transmission. The destination assembling-device receives the sub-transmissions sent byservers399a,399c,399(N−1) and399(N). It gathers573 the first fragment from each sub-transmission to reconstruct thefirst segment101a. In a similar fashion, additional fragments belonging to subsequent segments are obtained from the sub-transmissions, and used to reconstruct the segments. It is noted that any combination of sub-transmissions may be used, as long as a decodable set of fragments is obtained per each segment. It is also noted thatFIG. 6 illustrates a non-limiting embodiment and a sub-transmission may include two or more unique erasure-coded fragments per segment.
• In one embodiment, the push sub-transmissions is synchronous (all servers sending the fragments of each segment at approximately the same time). In another embodiment, the push sub-transmission is asynchronous and the arrival of different fragments associated with a specific segment at the assembling device side may be spread over a long period. This may occur, as an example, when some push servers are faster than others. In one embodiment using asynchronous sub-transmissions, the assembling device aggregates whatever fragments it can before presentation time of each segment, and then optionally supplements fragments using a pull retrieval process. A server that does not send fragments fast enough, and therefore usually causes supplemental requests, may be ordered to stop the sub-transmission. Another server may be requested, optionally by the assembling device, to replace the slow server by initiating a new sub-transmission.
• In one embodiment, the push-transmissions carry more erasure-coded fragments than needed for segment reconstruction. In one embodiment, the push transmissions carry fewer erasure-coded fragments than needed for segment reconstruction, and the remaining fragments are pulled by the assembling device.
• FIG. 8 illustrates one example of a fractional-storagesystem comprising servers699ato699(N) having abandwidth capability681. In other words, no server can send data at a rate higher than681. Assemblingdevice661 can select from which servers to obtain erasure-coded fragments for reconstruction of a segment. In one example, each server stores one relevant, unique, erasure-coded fragment. Therefore, from the N servers storing N possible unique fragments, the assembling device needs only K erasure-coded fragments for complete reconstruction of the segment (K<N). Since it is not important which K fragments from the N are retrieved, the assembling device may retrieve from the least loaded servers, so as to keep the load between the different servers balanced. When many assembling devices assemble contents in parallel, and since all assembling devices can select the least loaded servers, the end effect is that the load on the servers is balanced, with the potential for most servers to approach their maximal bandwidth capabilities. Optionally, that load balancing is achieved without significant coordination between the servers.
• In the example ofFIG. 8, assuming that K=3, the assemblingdevice661 may selectservers699b,699(N−1), and699afor fragment retrieval, as they have the lowest load of all N servers.Servers699cand699(N), as an example, will not be chosen, as they have relatively higher loads.
• The assembling device may select the least loaded servers using any appropriate method, such as, but not limited to (i) accessing a central control server having data about the load conditions on the various servers, or (ii) periodically querying the various servers on their load conditions.
• In one embodiment, instead of, or in addition to, selecting the least loaded servers, the assemblingdevice661 tries a random set of K servers from the N, and retrieves erasure-coded fragments from all servers reporting a load below a threshold, while higher loaded servers will be replaced by least loaded servers from the possible N servers. The end result is that the server array is balanced because the K erasure-coded fragments are retrieved from servers loaded below the threshold.
• In one embodiment, the assembling device does not know which of the servers store erasure-coded fragments related to the content to be retrieved, but the assembling device knows over how many servers (from the total number) the erasure-coded fragments are distributed. Therefore, the assembling device compensates for the infertile requests by enlarging the number of requests for erasure-coded fragments. Optionally, the requested servers are selected based on approximately random algorithm.
• FIG. 9 illustrates one embodiment ofdifferent servers698ato698(N) having different bandwidth capabilities of683ato683(N) correspondingly. Assemblingdevice661 selects from which K servers, out of the possible N, to retrieve the fragments for segment reconstruction, wherein each server may have different unutilized bandwidth and different bandwidth capability. When many assembling devices assemble contents in parallel, while rejecting servers with a high load, the end effect is that the server array is approximately balanced and most servers can approach their maximal bandwidth capabilities. In one embodiment, the server array is balanced by enabling many assembling devices to select the least loaded servers.
• Still referring toFIG. 9, in the example, and assuming that K=3,servers698a,698(N−1) and698(N) will be selected, as they have the highest unutilized bandwidth. In another example, the servers having the highest percentage of unutilized bandwidth will be selected.
• In one embodiment,servers698ato698(N) represent completely different types of server hardware, operating systems and capabilities, all put together in an array, and achieving load balance without the need for significant inter-server coordination. In one example, the fragments are distributed to at least two different classes of servers; the first class comprises high bandwidth CDN servers directly connected to the Internet backbone, and the second class comprises lower bandwidth CDN servers not directly connected to the Internet backbone.
• In one embodiment, it is desired to replace one or more servers by other servers for the delivery of erasure-coded fragments, wherein the replacement servers are selected using a second criterion from a pool of servers identified using a first criterion. For example, the first criterion for identifying the pool of replacement servers comprises looking for servers capable of increasing their fragment delivery throughputs, and the second criterion for selecting the replacement servers from the pool comprises selecting the best latency response server from the pool. In one example, the first criterion is a latency criterion, and the second criterion is a load criterion. In another example, the first criterion is a latency criterion, and the second criterion is a latency variance criterion. In another example, the second criterion is an approximately random selection. In one embodiment, a server selected using the second criterion is compared to the server to be replaced based on the second criterion. For example, the second criterion is latency, and the replacing server, selected from the pool, has a smaller latency than the server it replaces.
• In one embodiment, the server to be replaced is identified by comparing the actual performance level of the server with a threshold performance level. For example, when the compared performance is latency, a server having response latency above a certain threshold is replaced. In another example, the compared performance is the load on the server, which may be measured in terms of the amount of the unutilized fragment delivery bandwidth, or in terms of the percent of the server's unutilized fragment delivery bandwidth, or measured by any other appropriate technique.
• In some embodiments, the assembling devices use a fragment pull protocol to retrieve the fragments and approach the servicing servers. In some embodiments, the assembling devices use a push protocol to obtain the fragments and approach the servicing servers, possibly by obtaining multiple sub-transmissions comprising fragment sequences.
• In one embodiment, the servers are selected for fragment retrieval according to their unutilized fragment delivery bandwidth. For example, the servers report their unutilized bandwidth, and the assembling devices, or a control server, obtain the report and decide which servers to use for fragment delivery based on the unutilized bandwidth of each server.
• In one embodiment, the servers are selected for fragment retrieval according to their ability to support additional fragment delivery load. For example, the servers report their ability to support additional fragment delivery loads. And the assembling devices, or a control server, obtain the report, and select the servers that report an ability to support additional fragment delivery loads.
• In one embodiment, the assembling device, or a control server, looks for a pool of servers that may be used as replacements for servers that are loaded to a degree that does not allow continuation of fragment delivery. For example, the assembling device looks for potential unloaded servers, while retrieving fragments from other servers. The assembling device may sample relevant servers approximately randomly, and/or according to indications from a control server. The sampling process may comprise querying the potential server for load information, or measuring the latency or latency variance to the servers in order to estimate the current load on the server.
• It is noted that in some of the disclosed embodiments the storage gain of each segment on each server and the amount of fragments obtained from each server by each assembling device may be independent of the bandwidth capability of the server, and that the ability to load balance servers with different bandwidth capabilities/availabilities may not require adjustments of either storage gain or the number of fragments obtained from each server.
• In one embodiment, the amount of unutilized fragment delivery bandwidth, or the fragment delivery bandwidth capability of at least some of the servers changes on-the-fly. The assembling devices compensate for the changes approximately in real time by obtaining fragments from servers having enough unutilized bandwidth to provide the requested fragments.
• In one embodiment, a server that is loaded to a degree requiring replacement indicates so by responding to a fragment request, or a fragment load query.
• In one embodiment, a plurality of unsynchronized retrieving assembling devices, which optionally use fragment pull protocol, choose the least loaded servers from which to retrieve the erasure-coded fragments. Optionally, the servers have almost no inter-communication between them and the load balancing calculation is performed by the retrieving assembling devices. Because the assembling devices can select the least loaded servers, the assembling devices manage the load balancing. When the erasure-coded fragments stored by the servers are unique erasure-coded fragments, the retrieving assembling device may retrieve erasure-coded fragments from any relevant server. Therefore, it may be enough for the retrieving assembling device to have indication of the load on its targeted servers, and retrieve enough erasure-coded fragments from the least loaded servers.
• In one embodiment, a server signals the retrieving assembling device that it is close to its bandwidth limit and the assembling device searches for an alternative server. Optionally, the assembling device selects the server according to one or more of the following parameters: locality, cost, latency, or reliability. In one embodiment, the servers register their loads on a central server, and the assembling device selects the server to retrieve from, from the registered servers. In one embodiment, a central server, holding the loads of the various servers, determines for the assembling devices from which server to retrieve the erasure-coded fragments.
• In one embodiment, fractional-storage servers, from which the fragments are obtained for reconstructing a segment, are selected based on an approximately random selection algorithm from all of the servers storing the relevant fragments. In one example, an approximately random selection algorithm weighted according to the unutilized bandwidth of the servers is used for the approximately random selection of servers. The weighted random selection algorithm assigns servers with selection probabilities proportional to the amount of unutilized bandwidth for fragment delivery in each of the servers, such that the probability to select a server having a larger amount of unutilized bandwidth is higher than the probability to select a server having a lower amount of unutilized bandwidth.
• FIG. 10 illustrates one embodiment of a fractional-storage system. Assemblingdevice group661gobtain erasure-coded fragments from the servers, such that the resulting outgoing bandwidth utilizations of each server in the array is682ato682(N) correspondingly.FIG. 11 illustrates a case whereserver698bhas failed, itsbandwidth capability682b1 is zero, and is therefore unable to provide erasure-coded fragments. The assembling devices fromgroup661g, which previously obtained fragments fromserver698b, may attempt to access it again for additional fragments, but are now unable to get a response. These assembling devices therefore obtain fragments from alternative servers. The end effect is thatbandwidth682bis now loaded on the still available servers, such that thetotal bandwidth682a1 to682(N)1 approximately increases by a total amount equal to682b, optionally with no inter-server coordination, and simply by the fact that each assembling device selects alternative available servers for obtaining fragment on-the-fly. In one example, instead of obtaining fromserver682b1, the assembling devices obtain from the least loaded available servers. In one embodiment, a control server selects the alternative server/s for the assembling devices. In one embodiment, the assembling devices use a fragment pull protocol to obtain the fragments, and approach the alternative servers. In one embodiment, the assembling devices use a push protocol to obtain the fragments, and approach alternative servers, possibly by obtaining multiple sub-transmissions comprising fragment sequences. In this case, the sub-transmissions of the faulty server are discontinued and compensated for by other sub-transmissions from the alternative servers.
• FIG. 12 illustrates an example similar toFIG. 11 with the difference thatservers698a,698b, and698cto698(N) reside within, or get serviced via, first, second, and thirdInternet backbone providers300j,300i, and300hcorrespondingly. The group of assemblingdevices661gis connected to the Internet vianetwork300k, which has access to all three backbones, such that communication between the assembling devices andservers698ato698(N) pass via at least one of the backbones, or more. Ifserver698bis made unavailable to the assembling devices, optionally not due to a server failure, but rather due to congestion or a failure of the secondInternet backbone provider300i, assemblingdevices661gcompensate for the lost bandwidth by switching to the available servers on-the-fly. In one embodiment,networks300h,300i, and300j, are different physical sub-nets of one network connected to the Internet. In one embodiment, the assembling devices are connected tonetworks300h,300i, and300j, vianetwork300k, and then via one or more Internet Exchange Points (“IX/IXP”).
• In one embodiment, the assembling devices quickly react to the fault by approaching other responsive servers. In one example, the assembling devices use a fragment pull protocol to retrieve alternative fragments from responsive servers within less than 5 seconds of the fault event, and approximately without affecting the ongoing streaming operation.
• Referring again toFIG. 11, in oneembodiment server698bfails and a replacement is needed. The replacing server (not illustrated) may store either the same erasure-coded fragments stored onserver698b, or store other unique erasure-coded fragments associated with the segments stored on698b. One method for regenerating the erasure-coded fragments stored onserver698b, or generating equivalent unique erasure-coded fragments, includes the following steps: (i) identifying a failed/non-responsive server. (ii) determining the segments corresponding to the erasure-coded fragments that were stored on the non-responsive server. This may be achieved either by a query to a control server, or by a query to servers in the distributed storage that service the same contents. (iii) reconstructing each segment whose erasure-coded fragments are to be regenerated. This may be achieved by retrieving and decoding enough erasure-coded fragments. (iv) re-encoding at least the required erasure-coded fragments from the reconstructed segments. The re-encoded fragments may be the same as the erasure-coded fragments previously stored onnon-responsive server698b, or may be new, unique erasure-coded fragments. And (v) distributing the required erasure-coded fragments to a new replacement server for698b.
• The following embodiments describe processes for on-the-fly selection and re-selection of fractional-storage servers from which to obtain erasure-coded fragments.
• In one embodiment, a method for selecting enough new servers from which to obtain fragments, based on the unutilized bandwidth of the servers, includes the following steps: (i) accessing data regarding servers storing relevant fragments (referred to as the relevant servers); (ii) accessing data regarding the unutilized bandwidth of the relevant servers. Optionally, the data is received by the assembling device from the relevant servers; and (iii) obtaining fragments from enough of the relevant servers having approximately the highest unutilized bandwidth; or obtaining fragments from enough of the relevant servers selected randomly and having unutilized bandwidth above a certain threshold.
• In one embodiment, a method for selecting enough new servers from which to obtain fragments, based on locality, includes the following steps: (i) accessing data regarding the relevant servers; (ii) accessing data regarding the network topology distance (locality) from the relevant servers to the assembling device; and (iii) obtaining fragments from enough of the topologically closest relevant servers; or obtaining fragments from enough of the relevant servers that are located in the same sub-network as the assembling device, or located in the closest sub-networks.
• In one embodiment, a method for selecting enough new servers from which to obtain fragments, based on bandwidth and locality, includes the following steps: (i) accessing data regarding the relevant servers; (ii) accessing data regarding the unutilized bandwidth of the relevant servers; (iii) identifying more than enough relevant servers having the most unutilized bandwidth; or randomly identifying more than enough relevant servers having unutilized bandwidth above a certain threshold; (iv) accessing data regarding the network topology distance from the relevant servers to the assembling device; and (v) obtaining fragments from enough of the topologically closest relevant servers; or obtaining fragments from enough of the relevant servers that are located in the same sub-network as the assembling device, or located in the closest sub-networks.
• In the claims, a sentence such as “erasure-coded fragments encoded with a redundancy factor R>1 and associated with segments of streaming contents” is to be interpreted as erasure-coded fragments encoded with one redundancy factor or with a plurality of redundancy factors greater than one. For example, some fragments associated with a first set of segments of content may have a redundancy factor of two, and some fragments associated with a second set of segments of the same content may have a redundancy factor of three.
• In some embodiments, approximately random selection of fractional-storage servers is utilized for dealing with changes in network conditions, such as packets loss and/or server failure, without affecting the user experience, and optionally without prior knowledge of the type of the change in network condition. Optionally, new erasure-coded fragments are requested from the randomly selected servers instead of failed requests. Optionally, failed servers are replaced with other servers. Optionally, the combination and/or the number of fractional-storage servers from which the fragments are obtained changes over time. Optionally, the number of redundant fragment requests changes over time.
• In one example, a constant packet loss condition causes a constant fragment loss condition, which means that a certain percentage of fragments fail to be obtained by the assembling device. In this case, an approximately random selection of new servers may solve the problem, not necessarily because of the randomness of the selection (a general fragment loss condition may affect all servers), but simply because it generates more fragment requests to compensate for the loss, resulting in an increased fragment-delivery throughput that approximately levels at an average steady state value of: (Nominal_Throughput/(1−Fragment_Loss_Ratio)), wherein the Nominal_Throughput is the fragment-delivery throughput resulting when no packets are lost, and the Fragment_Loss_Ratio is the (fragment_lost/fragments_sent) ratio, which is a parameter that increases monotonically with the packet-loss. In another example, the failure is specific to one or more servers, and the approximately random selection of new servers finds new servers having lower failure ratios. In this case, the random selection solves the problem, since trying to retrieve again from problematic servers may have no positive effect. The above two examples demonstrate how a single selection strategy successfully copes with different types of failures, while resulting in a different behavior according to the type of failure (different resulting fragment delivery rates for example), and all that without prior knowledge of the exact nature of the failure. In another example, the servers are deployed over multiple networks and the communication fault comprises a failure of one of the networks causing related servers to be inaccessible. As a solution, the assembling device approximately randomly reselects the servers until it communicates with enough accessible servers to reconstruct a segment. Other examples are possible, in which an unknown failure is correctly handled by approximately random on-the-fly server selection.
• In one embodiment, different servers receive different weights proportional to their bandwidth. For example, the higher the bandwidth capability of the server, the higher the server coefficient; the higher the server coefficient, the higher the probability of selecting the server by an assembling device. In one embodiment, selecting the servers approximately randomly enables the fractional-storage system to operate well when the assembling devices do not know the load on at least some of the servers.
• In one embodiment, the approximately random selection of servers produces a set of source servers from which erasure-coded fragments are retrieved using a fragment pull protocol. In another embodiment, the approximately random selection of servers produces a set of source servers from which erasure-coded fragments are retrieved using a push-protocol. In this case, multiple sub-transmissions may be used to transport the fragments from multiple servers to an assembling device. When new server sources are randomly selected instead of others, the assembling device may end the sub-transmissions associated with the replaced servers, and initiate new sub-transmissions from the replacing servers, optionally from the point that the terminated sub-transmissions were interrupted.
• In one embodiment, the approximately random server selections are made from the servers not currently servicing the assembling device. In one embodiment, the approximately random server selections are made from all servers storing relevant fragments, including the server(s) that serviced the assembling device before being identified as problematic.
• In one embodiment, approximately random reselections of servers are performed occasionally, even if all currently servicing servers are functioning correctly. In this case, the assembling device may select a few servers from the current set, to be randomly replaced. In one embodiment, functioning servers are kept throughout several segment retrieval cycles, and potentially for the entire delivery cycle of a segmented content.
• In one embodiment, the assembling device does not know which of the servers store erasure-coded fragments related to the content to be retrieved, but the assembling device does know over how many servers (from the total number) the erasure-coded fragments are distributed. Therefore, the assembling device compensates for the infertile requests by increasing the number of requests for erasure-coded fragments. Optionally, the requested servers are selected based on approximately random algorithm.
• In one embodiment, a method for reselecting one or more fractional-storage CDN servers on-the-fly, comprising: pulling erasure-coded fragments from the servers; estimating the servers' load, latency, network congestion, and packet loss; and operating a fuzzy algorithm on the estimations in order to replace at least one of the servers with at least one other fractional-storage server. Optionally, the method further comprising operating the fuzzy algorithm based on measurements of many assembling devices and recommendations received from a central server. Optionally, the method further comprising replacing the servers quickly after missing a fragment. And optionally, the fuzzy algorithm weighs many possible solutions and converges to a sufficient one.
• In one embodiment, the fractional-storage system is approximately insensitive to the mixture of the consumed contents as long as the aggregated throughput is below the total throughput of the fractional-storage servers.
• FIG. 13 illustrates one example of a fractional-storage server array, including N servers (399ato399(N)), and storing content A, which includes erasure-codedfragments310ato310(N), and content B, which includes erasure-codedfragments320ato320(N). Each server is connected to thenetwork300 with a fragment deliverybandwidth capability B339. Therefore, the N servers have an aggregated bandwidth of B×N. A first group of assemblingdevices329aconsumes content A at anaverage bandwidth Ba349a. A second group of assemblingdevices329bconsumes content B at anaverage bandwidth Bb349b. Since all of the servers participate in the transmission of the two contents, the first and second groups can potentially consume all server bandwidth, up to the limit where Ba+Bb=N×B, with any ratio of demand between the first and second contents, and with no special provisions to be made when storing the erasure-coded fragments related to the two contents in the fractional-storage server array.
• FIG. 14 illustrates the case where thefirst group328a, which consumes content A, becomes larger than329a, with alarger bandwidth Ba348a. Thesecond group328b, which consumes content B, becomes smaller than329b, with asmaller bandwidth Bb348b, such that Ba is about the same as Bb. In this case, the array can still be exploited up to the aggregated bandwidth, since, as before, Ba+Bb can still be almost as high as N×B.FIG. 15 illustrates the case where the first group has disappeared, allowing thesecond group327b, which consumes content B, to extract an aggregated bandwidth ofBb347bthat can potentially reach the limits of the server array, such that Bb=N×B. Again, this is achieved without updating the erasure-coded fragments associated with content A and content B, and without using inter-server interaction.
• In some embodiments, the ability to utilize the aggregated bandwidth of approximately all of the participating servers, for the delivery of about any mixture of contents with about any mixture of content bandwidth demand, is made possible by one or more of the following: (i) each assembling device selecting a subgroup of the least loaded fractional-storage servers from which to retrieve the necessary number of erasure-coded fragments to reconstruct a segment or several segments (least-loaded server selection criterion); or (ii) each assembling device approximately randomly selecting a subgroup from which to reconstruct a segment or several segments, such that when many assembling devices select at random, the various fractional-storage servers are selected approximately the same number of times (or in proportion to their available resources, such as unutilized bandwidth), which in turn balances the load between the participating servers (random server selection criterion). It is noted that (i) the selections may be made by either the assembling devices themselves, or may be made for the assembling devices by a control server, which then communicates the selections to each of the assembling devices; (ii) the selections may be made approximately for each segment, or for a group of segments, or only once per content at the beginning of the content; (iii) some assembling devices may use an approximately random server selection criterion, while other assembling devices may use least-loaded server selection criterion; (iv) the least-loaded selected servers may be selected out of a portion of all available fractional-storage servers. For example, the least-loaded servers may be selected from fractional-storage servers with low latency response or with low hop count to the assembling device; (v) the least-loaded servers may include servers having the most unutilized bandwidth. Additionally or alternatively, it may include servers having any unutilized bandwidth left to serve additional assembling devices; (vi) an approximately random or least-loaded selection of servers may be made such that all servers are selected to determine a subgroup, or it can be made such that every time selections are made, only some servers are selected, while the others remain as before. In these cases, the assembling device runs a process in which only a small portion of the servers currently in the serving subgroup are reselected. In the case of approximately random selection, the assembling device may randomly select the number of servers in the serving subgroup for random selection (reselection in this case, since they are replacing other servers already in the serving subgroup of the specific assembling device), such that eventually, over time, all servers within the serving subgroup have the chance to be randomly reselected. In the case of least-loaded server selection, only the most loaded servers within the serving subgroup may be selected and replaced by less-loaded servers.
• In some embodiments, a broadcast-like effect is achieved by distributing to and retrieving from fractional-storage servers a broadcast channel/live content in real time, using a combination of real time distribution and real time retrieval techniques. In a broadcast-like effect, a given channel or content for broadcasting is distributed to at least one assembling device, optionally by means of pushing relevant fragments to the assembling device, or by pulling the relevant fragments by the assembling device, and potentially to many assembling devices at approximately the same time, which creates a similar effect to traditional broadcasting.
• FIG. 16 illustrates one embodiment of using the entire aggregated bandwidth of the fractional-storage servers for delivering multiple contents. Approximately any number of contents having any mixture of bandwidth demand per content may be delivered, as long as the aggregated bandwidth demand does not exceed the aggregated bandwidth of the fractional-storage servers. In one example, broadcast-like streams3101,3102, and3103 are delivered to multiple assembling devices via multiple fractional-storage servers. Each stream is a live TV channel carrying multiple TV programs. For example,stream3101 comprisesTV programs3110 to3112, each spanning a specific time interval. The other streams comprise of multiple TV programs as well. Before time T1,stream3130 has a bandwidth demand of3130′ (meaning that all assembling devices that are currently retrievingstream3130′ use a total bandwidth of3130′ out of the fractional-storage servers). Theother streams3120 and3110 have bandwidth demands of3120′ and3110′ respectively. The total bandwidth demand of the threestreams3130′+3120′+3110′ does not exceed the aggregated bandwidth of the fractional-storage servers3150, and therefore all streams are fully delivered to the assembling devices. The load of the three streams is spread approximately equally among the participating fractional-storage servers, optionally because of a mechanism that selects the least-loaded servers to serve each assembling device, and/or a mechanism that approximately randomly selects servers to serve each assembling device. At time T1,TV program3120 ends, andTV program3121 starts.Program3121'sdemand3121′ is higher than theprevious demand3120′, and therefore a higher aggregated bandwidth is drawn from the fractional-storage servers. Still, the aggregated bandwidth demand of all three streams (3130′+3121′+3110′) is lower than the maximum possible3150, and therefore the newly added bandwidth demand is fully supported by the servers. Optionally, the additional demand created by TV program3121 (3121′ minus3120′) is caused by the addition of new assembling devices that joinstream3102 and retrieving additional erasure-coded fragments. Additionally or alternatively, the additional demand created byTV program3121 is caused by a higher bandwidth demand ofTV program3121, such as 3D data or higher resolution. Newly added assembling devices may choose fractional-storage servers from which to retrieve, according to a least-loaded server selection criterion and/or an approximately random server selection criterion, and therefore the total load is still spread approximately equally among the participating servers. At time T2,TV program3110 ends, and anew program3111 begins, which is less popular, and therefore creates alower bandwidth demand3111′. The result is a decrease in the total delivered bandwidth. At timeT3 TV program3130 ends, andTV program3131 starts with a higher bandwidth demand of3131′. At time T4 bothTV programs3111 and3121 end, and twonew programs3112 and3122 start.TV program3112 is highly popular and therefore generates alarge bandwidth demand3112′.Program3122 is not popular, and therefore generates alimited bandwidth demand3122′. Some of the additional bandwidth needed byprogram3112 is taken from servers that stop serving assembling devices previously retrievingprogram3121, such that the aggregated bandwidth of all three streams (3131′+3122′+3112′) is still below the maximumpossible bandwidth3150, despite the fact thatprogram3112 is generating a large bandwidth demand. This example illustrates how the fractional-storage servers support almost any demand mixture, as long as the aggregated demand of all streams is kept below the aggregated maximum capacity of theservers3150. Consequently, the distribution of all of the streams to the fractional-storage servers is approximately unrelated to the changes in bandwidth demand for programs carried by each stream; each stream can be regarded as a sequence that is segmented, erasure-encoded, and distributed to the participating servers. There is no need to account for demand variations during the distribution of each stream, nor is there a need to know in advance the bandwidth demand for each stream or for each program within each stream. It is noted that the demand variations are illustrated as instant variations, but may also be gradual and may occur during a program and not necessarily when one program ends and the other begins.
• FIG. 17 illustrates one embodiment of a fragment pull protocol. Assembling device861 (also represented byprotocol diagram element810b) obtains erasure-coded fragments from fractional-storage servers899ato899(N) (also represented by protocol diagram element898), utilizing the following steps: (i) deciding810awhich segment to retrieve; (ii)device861 sending requests to some of the fractional-storage servers for erasure-coded fragments associated with the desired segment. For example, requests880ato880(K) for erasure-codedfragments890ato890(K), from servers899(a) to899(K), correspondingly; and (iii) the servers respond by sending the requested erasure-coded fragments. For example,servers899ato899(K) send881ato881(K) erasure-codedfragments890ato890(K) todevice861. The fragment request and receipt process begins at T1cand ends at T1d. At time T1d,device861 has enough erasure-coded fragments (K) to reconstruct the segment selected at810a. In one embodiment, the process from T1cto T1doccurs in real time, in support of streaming content presentation.
• FIG. 18 illustrates a similar process toFIG. 17, whererequest890bfails to result in a reception of erasure-codedfragment890bfor any reason (such as a server fault, network congestion, or abnormal latency conditions). Assemblingdevice861 therefore issues another request882(K+1) for erasure-coded fragment890(K+1) in response, and receives883(K+1) the additional erasure-coded fragment890(K+1) needed to reconstruct the segment.
• FIG. 19 illustrates a similar process toFIG. 17, where one or more extra erasure-coded fragments (in addition to the needed K) are requested in advance (illustrated as request880(K+1) for erasure-coded fragment890(K+1)), such that if, as an example, request890bfails to result in a reception of erasure-codedfragment890b, assemblingdevice861 does not have to request new erasure-coded fragments to reconstruct the segment, since there are still at least K erasure-coded fragments that were successfully received and therefore the segment can be reconstructed.
• FIG. 20 illustrates a similar process toFIG. 17, where requests for erasure-coded fragments are loaded into one aggregatedrequest870, that is sent to one of the fractional-storage servers (the receiving server is illustrated asprotocol diagram element888a, and will be also referred to as a “relay server”). In one example, if the relay server is899(N), then, it will forward the request toadditional servers899ato899c(protocol element888b) vianew requests870ato870c(on behalf of assembling device861).Servers899ato899cwill then respond by sending the erasure-codedfragments890ato890c(871ato871c) to theassembling device861. Server899(N) will send871(N) fragment890(N) to the assembling device.
• In one embodiment, an assembling device may aggregate several fragment requests into one message. The aggregated message is then sent to a fractional-storage server, possibly in a payload of a single packet, and optionally in order to conserve outgoing bandwidth and/or to reduce the number of packets needed to convey the requests. The fractional-storage server may then read the aggregated message and act accordingly by sending a plurality of fragment responses to the assembling device. The fragment responses may include one fragment at each payload, as is the case of responding to a single fragment request, or it may include an aggregated response including multiple fragments at each payload.
• In one embodiment, an assembling device may control the erasure-coded fragment reception throughput by controlling the rate of fragment request. For example, each of n fragments has a known size S1 to Sn. Therefore, issuing n requests over a period of T will result in an average fragment reception throughput of (S1+S2 . . . +Sn)/T. In one example, if each fragment is 1500 Bytes, and 64 fragment requests are issued over a period of 0.5 seconds, then the average expected fragment arrival throughput is (64×1500×8)/0.5=1.53 Mbps. The fragment requests do not need to be uniformly spread over the period of 0.5 seconds, although such a spread may result in a more stable throughput, which means that less communication buffering will be needed. Using the above-described rate-control technique may result in one or more of the following: retrieving the content at a target fragment reception throughput; preventing communication buffer spill at the last mile network resulting from uncontrolled fragment requests; and/or reducing fragment loss due to averaging the fragment traffic.
• The term “fragment pull protocol for high latency” as used herein denotes a protocol enabling an assembling device to request one or more fragments from one or more providing sources, wherein the time to transmit the one or more fragments in response to the assembling device request, through the slowest communication link connecting the responding source and the assembling device, is smaller than the round trip communication delay between the assembling device and the responding source, excluding the processing time of the providing source. For example, if the round trip communication delay between Israel and the USA is about 200 ms, the assembling device requests one fragment sized about 1500 bytes, and the slowest communication link is an ADSL line connecting the assembling device at 1.5 Mbps, then the time it takes to transmit the requested fragment through the slowest communication link is about 1500*8/1500000=8 ms, which is much smaller than the round trip delay. Many of the disclosed embodiments using fragment pull protocol may use fragment pull protocol for high latency for retrieving the fragments.
• In one embodiment, an assembling device transmits aggregated messages to a relay server, including the number of fragments needed per certain segment, but without identifying the storage servers from which fragments are to be requested. The relay server selects the appropriate storage servers to which the fragment requests are to be transmitted, and transmits discrete or aggregated fragment requests, corresponding to the number of fragments requested by the assembling device, to the selected storage servers. The storage servers receive the fragment requests from the relay server, and transmit the requested fragment to the assembling device. The relay server may select the storage servers according to one or more criteria, as long as the selected storage servers store relevant fragments. Optionally, the relay server forwards the address of the assembling device to the selected storage servers, and/or adds the address of the assembling device to the fragment requests transmitted to the selected servers, in order to enable the storage servers to transmit the fragment response to the assembling device.
• In one embodiment, shifting the process of selecting the storage servers from the assembling device to the relay server enables the design of a relatively thin and simple assembling device, having a relatively simple software, since all the assembling device has to decide in order to issue an aggregated fragment request to the relay server is how many fragments it needs per segment and, optionally, when it needs them.
• In one embodiment, an assembling device transmits aggregated messages to a relay server, comprising general information regarding a portion of streaming content for which fragments are needed. Optionally, the portion of the streaming content comprises several consecutive segments. In one embodiment, the portion is defined by a starting point and an ending point within the streaming content, and the relay server uses these points to determine the actual segments comprising the portion. Then the relay generates and transmits the corresponding fragment requests to the relevant storage servers.
• In some embodiments, the push protocol is implemented using one or more sub-transmissions. Optionally, a push protocol transmission is implemented using multiple sub-transmissions, each transporting a fraction of the fragments transmitted by the push protocol transmission. A sub-transmission may be transported using an IP stream such as RTP, an HTTPS session, or any other form of transporting a sequence of fragments between a source server and a destination assembling device.
• In one embodiment, an assembling device starts retrieving fragments using only fragment pull protocol processes, and then, when concluding that a specific server is responsive enough, instructs it to start sending a push-transmission for the remaining segments. In this case, the assembling device may start with pure pull-protocol based fragment retrieval, and gradually switch to push-protocol transmissions, up to the point that approximately all fragments are delivered using push-transmissions, and using the pull requests only as a means to overcome failure of obtaining specific fragments by the assembling device. In one embodiment, the fragment pull protocol and the push protocol are used interchangeably to obtain enough fragments to reconstruct segments. In one embodiment, the assembling device may start to obtain fragments using a push protocol and then switch to a fragment pull protocol. In one embodiment, the assembling device may use both fragment pull protocol and push protocol to obtain fragments at the same time, wherein the assembling device may change the ratio Fpull/Fpush on-the-fly to any value between zero and infinity, where Fpull denotes the number of fragments associated with a certain segment that are obtained using a fragment pull protocol, and Fpush denotes the number of fragments associated with the certain segment that are obtained using a push protocol.
• In the claims, sentences such as “wherein the assembling device is configured to use a fragment pull protocol to obtain the fragments” and “wherein the assembling device is configured to use sub-transmissions to obtain the fragments” are to be interpreted as open claim language. Therefore, an assembling device configured to use a fragment pull protocol to obtain fragments may also obtain fragments using sub-transmissions, and vice-versa.
• In one embodiment, different quantities of erasure-coded fragments are generated per different segments. In one embodiment, some segments store data that is considered more important than data stored in other segments, and relatively more erasure-coded fragments are generated from the segments storing the more important data than from the segments storing the less important data.
• In some embodiments, the content is segmented into a plurality of segments to enable beginning to play the content as it is being obtained, and optionally enable trick play. The different segments may or may not be of the same size.
• In one example, the content supports streaming presentation, and the segments are small enough to enable presentation shortly after beginning the reception of the first segment(s). For example, each segment may include 96 KByte, allowing a 5 Mbps receiver to download the segment in approximately 0.2 seconds, and optionally begin the presentation shortly thereafter. In one embodiment, the time to play is reduced by segmenting certain portions of the content into smaller segments, while the remaining portions are segmented into larger segments. A smaller segment can be retrieved faster, while a larger segment may be better optimized for storage gain and/or efficient transmission.
• In one embodiment, the short segments are 96 Kbytes in size, and the long segments are 960 Kbytes in size. The redundancy factors used for encoding short and long segments into fragments are 100 and 5 respectively. 1500 Bytes fragments are used for both sizes. The short segments are therefore encoded into (96K/1500)×100=6,400 fragments, from which only about 64 are needed for reconstruction, and the long segments are encoded into (960K/1500)×5=3,200 fragments, from which only about 640 are needed for reconstruction. Short segments are reconstructed more quickly than long ones, as they require fewer fragments to be decoded. Optionally, each fragment is stored on a different server, resulting in a storage gain of 64 for short segments, and 640 for long segments.
• The following embodiments discuss different methods for segmenting the content. In one embodiment, at least one portion of the content is segmented into multiple segments in sizes within a first size range, and the remainder of the content is segmented into a plurality of segments in sizes within a second size range (additional size/s may be added similarly). The sizes included in the second size are larger than the sizes included in the first size range. Pluralities of erasure-coded fragments are generated from each of the segments. The segments of sizes within the first size range are better suited for fast retrieval, and the segments of sizes within the second size range are better suited for high-gain storage. In one example, the segments in sizes within the first size range belong to approximately the beginning of the content. In one example, the segments in sizes within the first size range belong to locations within the content requiring trick play access. In one embodiment, the segments of the first type are encoded into fewer fragments than the segments of the second type. This allows a fast retrieval of the shorter segments.
• In one example, thecontent100 is a 1 GByte encoded H.264 file, storing a 2-hour motion picture, and is segmented into approximately 10,000 segments of approximately 100 Kbytes each. In another example, thecontent100 is a 4 MByte web-site information (HTML, FLASH, or any other combination of information that encodes the presentation of a website), and is segmented into 4 segments of approximately 1 MByte each.
• FIG. 21 illustrates one example in which thecontent100 is segmented into segments, such that thefirst segment104ais smaller than theconsecutive segment104b, which is smaller than followingsegments104cand104d. In another example, thecontent100 is segmented into segments, such that the first several segments (e.g.104aaand104bb, which are the same size), are smaller than consecutive segments (e.g.104ccand104dd, which are the same size).
• FIG. 22 illustrates one example in which thecontent100 is segmented into cyclic sets of successive segments increasing in size. For example,105bis equal or larger in size than105a, and so on, up tosegment105d;105fis equal or larger in size than105e, and so on, up tosegment105h. In one example,segment105eis equal in size tosegment105a. Point105EP represents the ending of the first set, and the beginning of the second set.
• In one embodiment, segments are created on-the-fly, such as during a live event or when the content is made available to the segmentation process as an on-going stream. In one embodiment, the content supports streaming presentation, and the segments are of the small size, to enable content presentation shortly after beginning the reception of the first segment (or any other segment). In addition, the erasure-coded fragments are kept as small as possible, while still enabling efficient transport over an IP network. For example, each erasure-coded fragment is about 1500 Bytes and can be transported using one IP packet.
• It is to be noted that streaming content may also be manifested as an intermediate product of a process. For example, in a case where a video camera outputs erasure-coded fragments that can be decoded into streaming content, the intermediate data from which the erasure-coded fragments are generated is considered to be streaming content (even if the video camera does not output that intermediate data). Moreover, streaming content may include: content that is produced and then immediately transmitted to a receiving server, content that is produced but stored for any length of time before being transmitted to a receiving server, content that is transmitted to a receiving server and then immediately sent from the receiving server to a client, content that is transmitted to a receiving server, then buffered for some time at the receiving server and then sent from the receiving server to a client, content that is solely played at a client, and content that is manipulated or changed or reacted to at the client while a continuation of the content is still being played at the client.
• FIG. 23 illustrates one example of geographically distributed fractional-storage servers399ato399n, in whichservers399ato399care located inEurope676,servers399dto399gare located on the east coast of theUS677,servers399hto399iare located on the west coast of theUS678 andservers399kto399nare located inJapan679. Assembling devices all over the world obtain erasure-coded fragments from the globally distributed fractional-storage servers. The characteristics of the fractional-storage system, according to some embodiments, allow the globally distributed assembling devices to exploit the outgoing bandwidth of the globally distributed fractional-storage servers approximately up to the point where allservers399ato399nutilize their available outgoing bandwidth for content delivery.
• In one embodiment, the main demand for fragments shifts between the different global locations as the day elapses. For example, at 8 pm Pacific Standard Time, the main fragment demand is generated from the US west coast. At that time, the local time in the east coast is late evening, the time in Europe and Japan is early morning and noon respectively, and thus very little fragment demand is generated from these regions. The high fragment demand load generated from the west coast is spread across all of the fractional-storage servers. As the day elapses, the load generated from the west coast declines, and the main load shifts to Japan as time there becomes afternoon. When that happens, the servers are still able to supply all content demands, as they are still able to deliver maximal bandwidth to assembling devices in Japan. As the cycle continues, the main load shifts again from Japan to Europe, from Europe to the US east coast, and from there back to the US west coast, following a 24-hour cycle. In some embodiments, the servers are able to deliver maximal fragment traffic, resulting from peak demands occurring during a day cycle, to anywhere on the globe.
• In one example, there are 14 globally distributed fractional-storage servers; each server has a bandwidth of B, and the total capacity of the array is 14×B. Assuming the total global peak demand during the daily cycle does not exceed Bg, then the system is balanced and can meet all demands during the daily cycle if Bg<14×B, meaning that B>Bg/14. In this example, all servers may be at, or may approach, their peak bandwidth capabilities for a relatively long period, and feature relatively short idle periods. In one example, the number of servers in the global array is 10,000, from which 2,500 are located on the US west coast, 2,500 on the east coast, 2,500 in Europe and 2,500 in Japan. In one example, the number of servers in the global array is 1,000, from which 100 are located on the west coast, 700 on the east coast, 100 in Europe and 100 in Japan.
• In one embodiment, multiple contents originating from multiple global locations (and therefore expected to require high loads at different times of day), are all stored on the globally distributed fractional-storage servers. Therefore, the system's bandwidth capacity equals the aggregated bandwidth of its server members, optionally regardless of which content generates high load, regardless of when the load is generated during the day, and regardless of where the load is generated from.
• In one embodiment, at some point in time, some portions of the Internet may become congested at some global locations. The global system assures that servers not affected by the congestion handle the excess load, such that operation close to peak bandwidth performance is still possible.
• In one embodiment, the globally distributed assembling devices retrieve fragments from the fractional-storage servers using a fragment pull protocol, and determining which servers deliver fragments to which assembling devices load balances the distributed system. In one embodiment, the globally distributed assembling devices obtain fragments from fractional-storage servers using a push protocol with multiple sub-transmissions, and determining which servers deliver fragments via the sub-transmissions to which assembling devices load balances the distributed system.
• FIG. 24 illustrates one embodiment in which assembling devices distributed over different time zones together induce fragment traffic having a reduced peak-to-average traffic ratio, as compared to the fragment traffic induced by assembling devices located in any single time zone.Graph1602 illustrates the fragment traffic induced by assembling devices located at a first time zone. The peak ofgraph1602 occurs during the late afternoon, local time of the first time zone. Similarly,graphs1603 and1604 illustrate induced traffic from second and third time zones. Since the first, second and third time zones are different, the peak traffic of each graph occurs at a different time. The peak-to-average fragment traffic ratios ofgraphs1602 to1604 are relatively high, since most of the traffic is generated close to the peak demand. In the case of video traffic, a daily peak-to-average traffic ratio of about six is expected during one day, starting at T1 and ending at T2. The combined traffic induced by all assembling devices is the sum ofgraphs1602 to1604, which is schematically illustrated asgraph1601. Since the peaks ofgraphs1602 to1604 occur at different times, the combinedtraffic1601 behaves much more smoothly and has peaks close to the peaks ofgraphs1602 to1604, resulting in a much lower peak-to-average traffic ratio, which in some embodiments is about two or three. This means that the fractional-storage servers can be utilized during longer periods of the day when servicing assembling devices located at different time zones. In one embodiment, the distribution of the assembling devices to the different time zones results in an approximately flat traffic during the day, having a peak-to-average traffic ratio approaching one. Such a distribution is challenging in real life deployments, but can be approached by engineering the distribution of the assembling devices over the globe.
• Still referring toFIG. 24, in one embodiment, the severs are connected to the Internet using guaranteed fixed bandwidth communication links, and can together deliver to the Internet fragment traffic of1610 all day. In this case, it is clear thattraffic graph1601 utilizes the fixedbandwidth capacity1610 better than any of thegraphs1602 to1604, since it approaches the maximal capacity for longer periods over the day.
• In one embodiment, the servers are spread over two or more continents, and some of the fragments associated with the same segments are stored on different servers located on different continents. This achieves content placement diversity, and results in better immunity to different network and server faults.
• FIG. 25 illustrates one embodiment in which US-based fractional-storage servers399a′ to399n′ deliver erasure-coded fragments to assembling devices spread over the globe. The assembling devices spread over the globe induce a total fragment traffic from the US-based servers having a reduced peak-to-average traffic ratio, as compared to the fragment traffic induced by assembling devices located in any single time zone. In one example, 5,000 fractional-storage servers are located in the US and service 10 million assembling device subscribers spread over the globe. At a first period during the day, the servers deliver erasure-coded fragments concurrently to 2 million assembling devices located primarily in Japan. At a second period during the day, the servers deliver erasure-coded fragments concurrently to 2 million assembling devices located primarily in Europe. At a third period during the day, the servers deliver erasure-coded fragments concurrently to 2.5 million assembling devices located primarily on the East Coast, and ½ million assembling devices located primarily on the West Coast. At a fourth period during the day, the servers deliver erasure-coded fragments concurrently to ½ million assembling devices located primarily on the East Coast, and 2.5 million assembling devices located primarily on the West Coast. According to this example, the servers are capable of delivering a peak fragment traffic resulting from the demand of at least 3 million assembling devices concurrently.
• In one embodiment, the servers are spread over different time zones. Different servers located at different time zones usually encounter peak load conditions at different times, especially if they share resources, such as communication link to the Internet, processing resources, storage, Tier-1 ISP networks, backbone networks, or any other resources with local servers delivering general Internet traffic. Load conditions may refer to actual load on the servers, load on a communications link connecting the server to the Internet, load on a local backbone or Tier-1 network, or any type of condition in which additional fragment traffic will contribute to service degradation. In the case of a load condition, it is advantageous to refrain from obtaining fragments from servers that directly contribute to the load, and to try to obtain fragments from servers that do not directly contribute to the load. Servers encountering load conditions below a certain threshold are usually found somewhere, as they are spread over different time zones, and these servers are the preferred fragment sources.
• FIG. 26 illustrates one example of different loads at different times for different time zones.Graphs641a,641b,641cand641drepresent load levels encountered byserver groups679,676,677 and678 respectively, located in the Far East, Europe, the US east coast, and the US west coast respectively. In one example, the loads refer to traffic levels on communication links connecting the data centers, in which the servers are placed, to the Internet. In this case, the traffic may be general Internet traffic generated by servers and other application/s not necessarily related to fragment delivery, and the communication links can also be referred to as shared links, as they are used to transport both fragment traffic and general Internet traffic. During a 24-hour period, all encountered load levels complete one cycle. The load level graphs are shifted in time in respect to each other, according to the time shifts between the various time zones around the world in which the different server groups are located. As an example, graph641arepresents load encountered by the servers in the Far East, with a peak load occurring about 7 hours beforegraph641brepresenting load encountered by the servers in Europe.
• At each arbitrary point in time, server groups around the world may encounter different load conditions. As an example, atpoint642a,server group679 encounters medium load conditions,server group676 encounters peak load conditions, andserver groups677 and678 encounter low load conditions. Therefore, at the point intime642a, it is beneficial for assembling devices to obtain erasure-coded fragments only fromserver groups677,678, and maybe679.Server group676 encounters peak load conditions, and therefore will not be approached by the assembling devices. At a different point intime642b, the worldwide load conditions change, such thatserver groups679 and676 encounter low load conditions, andserver groups677 and678 encounter high load conditions. At this point, assembling devices will obtain fragments fromservers groups679 and676 and will refrain from approachingserver groups677 and678.
• In one embodiment, the load conditions encountered by each server group, or by specific servers, are published by the servers. In one embodiment, the load condition level encountered by each server is sent to each assembling device as a response to an erasure-coded fragment request.
• In one embodiment, the communication link transporting fragments from a server or group of servers to the Internet is owned by a data center operator. The data center operator publishes the load condition associated with the link. The published information is used to select servers that transmit fragments via relatively unloaded links as compared to other links.
• In one embodiment, the load conditions encountered by a server are detected by an outside source, such as an assembling device or a control server, using one of the following methods: (i) detecting an increased latency in responding to a request such as a fragment pull protocol request, (ii) detecting a certain level of latency variance, (iii) detecting a certain level of packet or fragment loss, and/or (iv) detecting outages in server's traffic.
• FIG. 27 illustrates one embodiment of data centers communicating via shared links. Fractional-storage servers1699ato1699care collocated with at least onegeneral server1698 in adata center1671. All the servers are connected to the Internet via a sharedcommunication link1681. Therefore, erasure-coded fragment traffic transmitted by the fractional-storage servers and general Internet traffic transmitted by the general server are mixed together on the sharedlink1681. Similarly, fractional-storage servers1699dto1699gare collocated with at least onegeneral server1699 in adata center1672, and share thesame communication link1682 to the Internet. In one embodiment, the fractional-storage servers are selected for fragment transmittal when the communication link through which they transmit fragments to the Internet is loaded below a certain level. This principle is demonstrated by the following example: assuming that any three fractional-storage servers out of1699ato1699gstore a decodable set of fragments, the three servers will be selected according to the load of the link through which they communicate. If thegeneral server1698 transmits a high level Internet traffic vialink1681, and this traffic is close to the maximum capacity of the link, then using any ofservers1699ato1699cis not advisable. Instead, in a case where thegeneral server1699 does not create a high level traffic andlink1682 is relatively free to transport fragments, any three servers out ofservers1699dto1699gmay be used. When the fractional-storage servers deliver fragments to many assembling devices, servers transmitting via relatively unloaded links are preferred, such that the end effect is thatservers1699dto1699gdeliver a higher fragment load thanservers1699ato1699c. In other words,servers1699dto1699gparticipate in more sub-sets of servers delivering decodable sets of fragments to assembling devices thanservers1699ato1699c.
• FIG. 28 illustrates one embodiment of alternative servers communicating via shared networks. Fractional-storage servers1699a′ to1699c′ transmit erasure-coded fragment traffic over Internet backbone networks or Tier-1networks1661 and1662. The fragment traffic and the general Internet traffic transported via the networks are mixed together on the networks. Similarly, fractional-storage servers1699d′ to1699g′ are connected to Internet backbone networks or Tier-1networks1663 and1664. In one embodiment, the fractional-storage servers are selected for fragment transmittal when the networks through which they transmit fragments to the Internet are loaded below a certain level. This principle is demonstrated by the following example: assuming that any three fractional-storage servers out of1699a′ to1699g′ store a decodable set of fragments, the three servers will be selected according to the load of the network through which they communicate. If the general Internet traffic transported vianetworks1661,1662 is close to the maximal capacity of the networks, then using any ofservers1699a′ to1699c′ is not advisable. Instead, in a case wherenetworks1663,1664 are relatively unloaded with general Internet traffic, any three servers out ofservers1699d′ to1699g′ may be used. When the fractional-storage servers deliver fragments to many assembling devices, servers transmitting via relatively unloaded networks are preferred, such that the end effect is thatservers1699d′ to1699g′ deliver a higher fragment throughput thanservers1699a′ to1699c′. In other words,servers1699d′ to1699g′ participate in more sub-sets of servers delivering decodable sets of fragments to assembling devices thanservers1699a′ to1699c′.
• In one embodiment, the traffic loads on the sharedlinks1681 and1682, or sharednetworks1661,1662 and1663,1664 change to below a first level and above a second level, and the servers are dynamically selected accordingly. In one embodiment, the changes in the traffic loads result from changes in local Internet traffic demands during a 24-hour cycle. Different servers are located in different time zones, such that the peak of the changing traffic load occurs at different times for different servers. Servers transmitting via relatively unloaded links or networks are preferred over servers transmitting via relatively loaded links or networks as the load cycle progresses. In one embodiment, the load changes below a first level and above a second level for different links or networks at different times, and the servers are selected accordingly. For example, only servers that communicate via links or networks loaded below the first level are approached by the assembling devices.
• In one embodiment, when the shared link or network is loaded below a first level, the number of sub-sets in which the servers accessed via the shared link or network are allowed to participate is increased in order to increase the fragment consumption from these servers. When the shared link is loaded beyond a second level, the number of sub-sets is decreased. In one example, the amount of fragment traffic transmitted by a server is directly coupled to the number of sub-sets in which the server participates.
• In one embodiment, the maximum number of sub-sets of servers delivering decodable fragments to assembling devices in which the servers accessed via the sharedlinks1681 and1682 or sharednetworks1661,1662 and1663,1664 are allowed to participate is approximately a decreasing function of the throughput of the general Internet traffic via the shared link or network. In one example, as the general traffic increases, the server participates in fewer sub-sets, and above a certain point the server does not participate in any of the sub-sets.
• In one embodiment, an assembling device will refrain from requesting fragments from a server encountering load conditions close to maximal load, or above a certain threshold. This mechanism may be used to lower the cost of placing a server or a virtual server in a colocation center or any other data center, as the geographically distributed fractional-storage servers do not consume bandwidth and/or processing resources during peak load periods. Furthermore, this mechanism may be used to lower the cost of Internet bandwidth connections to the geographically distributed fractional-storage servers, as the servers do not consume Internet bandwidth during peak load periods.
• In one embodiment, the selection of which fractional-storage servers deliver erasure-coded fragments to which assembling devices approximately determines the network paths through which the fragments are transported. When the system has a redundancy factor greater than 1, there is a degree of freedom in selecting the servers that can deliver a decodable set of fragments to an assembling device. If the servers are spread over different networks, then each server, or groups of servers, may have different networks path through which fragments flow when transmitted to an assembling device. Selecting the servers therefore means selecting network paths through which fragments are delivered to an assembling device. As the redundancy factor, the storage gain, and the diversity at which servers are spread over different networks increase, so does the number of potential network paths resulting from server selections. The selection of paths, via selection of servers, can be used to avoid congested networks, to prefer certain paths that are more cost effective, or to optimize any other criterion related to fragment flow paths.
• FIG. 29 toFIG. 31 illustrate the influence of selecting source servers on backbone traffic.FIG. 29 illustrates one example wherein fractional-storage servers3599ato3599jare grouped in threelocations3541,3542, and3543, connected to the Internet vianetworks3505,3402, and3509 respectively. Assemblingdevices3537,3538, and3539 are connected to the Internet and obtain fragments from the servers. Assuming any three servers can be used to deliver decodable sets of fragments to the assembling devices,servers3599a,3599d, and3599hare selected to deliver fragments to assemblingdevice3539. In this case, the resulting three network paths through which fragments flow to the assembling device are (i) fromserver3599a:first path3509,3501,3403 (ii) fromserver3599d:second path3505,3503,3501,3403, and (iii) fromserver3599h:third path3402,3508,3502,3501,3403.
• FIG. 30 illustrates one example whereinnetworks3502,3504, and3508 get congested with Internet traffic, not necessarily as a result of fragment traffic generated byservers3599ato3599j, and possibly as a result of general Internet traffic. The third path includes two of the congested networks:3508 and3502, and should therefore be avoided. This means that another server, instead of3599h, has to be selected, such that it does not result in a fragment deliverypath comprising networks3508 and3502.Server3599bis therefore selected, resulting in a fragment delivery path of3509,3501,3403, which is similar to the first path already delivering fragments fromserver3599a.Assembling device3538 will use theservers3599hto3599j, as they are the only servers that avoid the congested networks. The path in this case comprisesnetworks3402 and3401.Assembling device3537 can use any three of the servers belonging togroups3541 and3543.
• In one embodiment, the different networks are associated with different costs. The cost may be related to any of the following parameters, or other parameters relevant to transporting fragments over a network: (i) network's congestion level, (ii) network's remaining capacity, (iii) network's packet loss, (iv) network's latency, (v) network's latency variance, and/or (vi) the fee for transporting bits over the network. In one example, selecting which servers deliver fragments to which assembling devices is performed such that the resulting fragment delivery paths comprise networks having the least aggregated cost, or a competitive aggregated cost compared to alternative paths.FIG. 31 illustrates one example of assigning costs to network paths. Each of the networks is associated with a cost of 1 to 4. The higher the cost, the more congested the network.Assembling device3539 can obtain fragments from eitherserver group3541,3542, or3543. The resulting three paths have the following aggregated costs: (i) first path, from group3543: 4+1+1=6, (ii) second path, from group3541: 3+1+1+1=6, (iii) and third path, from group3542: 1+2+2+1+1=7. The servers are selected from the first and second groups, as the resulting path cost is 6. Servers from the third group are usually not selected, as the resulting path cost is 7.
• FIG. 32 illustrates one embodiment wherein the selection of which servers deliver fragments to which assembling devices is used to determine network paths for fragment delivery. The servers are selected such that the resulting paths: (i) avoid certain loaded routers, and/or (ii) comprise routers having an aggregated cost lower than other possible paths. Fragment traffic going from groups ofservers3541,3542,3543 to anassembling device3539 may pass through any of therouters3501 to3506, depending on which three servers are selected for fragment transmission. In one example,router3506 is congested. Therefore, only serves3599dto3599gand3599hto3599jare considered for fragment delivery, in order to avoid transporting the fragments via thecongested router3506.
• Network paths, networks, and/or servers, which should be avoided, may be identified using one or more of the following embodiments. In one embodiment, the operator/owner of the networks/routers indicates that certain networks/routers are to be avoided. In one embodiment, the networks/routers are associated with a cost that is used for selecting the paths. In one embodiment, the different paths are empirically checked by transporting traffic from servers to assembling devices, and measuring parameters such as latency, latency variance, fragment or packet loss, and/or traffic outages. In one embodiment, certain networks/routers are to be avoided during a certain period of the day, and can be used during another period of the day. For example, an Internet bandwidth provider has a high traffic load on one of its network links during the afternoon, but this same link is almost free of traffic during the early morning. In this case, the provider can indicate that fragments can be delivered via the link only during early mornings. In another example, an Internet backbone provider has a high traffic load on one of its Tier-1 networks during the evenings, and a moderate load during the noon period. In this case, the process of selecting the fragment delivering servers will consider this, and select delivery paths comprising the Tier-1 network only during the noon period.
• In one embodiment, after obtaining some data regarding some of the loads, availabilities, losses, costs, preferences, and/or any other data that may influence the selection of the servers, algorithms and/or theorems such as Minimax (also known as Minmax) may be used for optimizing the selections.
• In some embodiments, the path though which a fragment will flow from a server to an assembling device may be estimated using one or more of the following: (i) TraceRoute functions to map routers between the various servers and the assembling device, or (ii) obtaining a topological map of the Internet, and estimating the paths accordingly. The estimated path may then be used to shape the actual fragment flow paths by selecting fragment-delivering servers. In one embodiment, the path through which fragment flow is unknown, and the determination of which servers deliver fragments to which assembling devices is performed approximately randomly, until an indication is received that a certain network, or router, or groups of such, are avoided.
• By using a pull protocol or a push protocol with multiple sub-transmissions, the assembling device can obtain erasure-coded fragments from one, two or more different arrays of CDN servers and/or bandwidth amplification devices seamlessly.
• In one embodiment, when a CDN server receives a request for an erasure-coded fragment, it may supply the erasure-coded fragment or supply an address of a bandwidth amplification device having an image of the requested erasure-coded fragment. Optionally, a bandwidth amplification device storing one erasure-coded fragment of a specific content also stores an image of some or all other erasure-coded fragments associated with the specific content (which are stored on the specific CDN server). Alternatively, the bandwidth amplification device stores unique erasure-coded fragments generated from the same segments used for generating the erasure-coded fragments stored on the specific CDN server. In these cases, the assembling device may approach the bandwidth amplification devices instead of the CDN server for the relevant erasure-coded fragments of the specific content until (i) the end of the content; (ii) a predefined time period elapses; (iii) receiving an appropriate message; or (iv) a combination of the aforementioned.
• In one embodiment, an assembling device tries to obtain an erasure-coded fragment or sub-transmission from the relevant server, and if the server does not have the necessary bandwidth to respond with fragment/s, the server relays the fragment request/s to relevant bandwidth amplification devices. The relevant bandwidth amplification devices can then send the fragment/s directly to the assembling device.
• In one embodiment, unique erasure-coded fragments can be distributed between two types of devices: (i) high bandwidth fractional-storage servers, such as CDN servers, and (ii) relatively low bandwidth and storage devices acting as bandwidth amplification devices, such as peer-to-peer (P2P) devices. Since the fragments distributed between the two types of devices are unique, any combination of devices, from both types, can be used to obtain a decodable set of fragments, if the combination of devices stores a decodable set of fragments. In one embodiment, there are at least ten times more bandwidth amplification devices than high bandwidth servers, and the redundancy factor used in decoding the fragments is greater than 10. In this case, the servers can be used all or most of the time, and the bandwidth amplification devices can be used from time to time, according to bandwidth requirements, and according to the availability of the bandwidth amplification devices. In one embodiment, the processes of obtaining a fragment from a server and from a bandwidth amplification device are essentially the same, and the fragments are essentially identical in construction and format. In one embodiment, the high redundancy factor needed to support a large hybrid array of servers and bandwidth amplification devices is achieved using rateless coding techniques.
• FIG. 33 illustrates one embodiment in which fractional-storage servers3799ato3799cstore a first portion of rateless-coded fragments; and a large number of P2Pbandwidth amplification devices3799dto3799jstore a second portion of the rateless-coded fragments. Decodable sets of fragments can be obtained from combinations of fragments from the first and second portions of the fragments. Optionally, the fragments are obtained approximately only from P2P devices serviced byISPs3771,3772 having communication lines estimated not to be overloaded by additional fragment traffic.
• In one embodiment, the P2P devices are spread over different time zones spanning at least three hours, and the fragments are obtained approximately only from P2P devices located in time zones in which the current local Internet traffic is relatively low in comparison to peak local traffic.
• Referring again toFIG. 3 withdevice6610 as a non-assembling CPE, such as a STB, PC or gaming console, capable of performing standard request, reception, and decoding of video over IP network. In one embodiment,server661s—also referred to as proxy server, assembling server, and in some cases assembling device—performs three primary functions: (i) receipt of content requests fromnon-assembling client device6610; (ii) assembly of content, as requested by client661o, from the fractional-storage servers and optionally from the bandwidth amplification devices; (iii) optionally, conversion of the assembled content into a streaming format; and (iv) transmission of the streaming content to the requesting client661o.Client6610 can then store the content, or present it. In one embodiment, the assembled content is a general web content, including HTML, FLASH or any other data format that can be found in a web-based site.
• In one embodiment, althoughserver661sis illustrated as being connected to network300 on one side and to network300non the other,server661smay also be connected to another network element, such as a router, which makes the topological connection betweennetworks300 and300n. In that case,server661scommunicates with bothnetworks300 and300nvia the other network element.
• In one embodiment, a CDN is created by the aggregated bandwidth and storage capacity of the participating erasure-coded fractional-storage servers. In one example, a large scale CDN includes several hundreds or thousands of fractional-storage servers connected to the Internet. These servers send erasure-coded fragments to a large number, potentially millions, of assembling devices. In order to keep costs low for sending a large number of fragments from fractional-storage servers to assembling devices, the servers are located on the Internet backbone, or close to it.
• The current Internet backbone primarily comprises different Tier one ISP (or other) networks that interconnect at various Internet Exchange Points (IX or IXP), using peering agreements. Tier one ISPs, or other backbone-forming network entities, can reach any portion of the Internet via other Tier one ISPs or other backbone-forming networks, without paying any Internet transit fee, and solely by utilizing mutual peering agreements. In order to gain access to large amounts of inexpensive bandwidth, the fractional-storage servers are typically located on the Internet backbone. This means that the servers are either co-located (and connected) with a core switching router that interconnects the Internet backbone networks at an IXP, or, alternatively, co-located (and connected) with a router which is part of the backbone network, typically located at a data center or co-location center. Fractional-storage servers can also be located close to the Internet backbone, which means that they are co-located (and connected) with a router which is part of a Tier two ISP network, which has a high bandwidth connection with at least one Tier one operator, to which it pays transit fees in order to potentially reach all portions of the Internet.FIG. 34 illustrates one example of a fractional-storage server3001, which is one of a plurality of servers forming a large-scale CDN, located on the Internet backbone by being connected to the Internet backbone viaIXP 3091. In a second example, fractional-storage server3002 is located on the Internet backbone by being connected to a Tier onebackbone network3080. In a third example, fractional-storage server3011 is located close to the Internet backbone by being connected to a Tier twoISP network3070, which is connected to the backbone via Tier oneISP network3081. In one embodiment, a typical fractional-storage server is located on the backbone or close to the backbone by being attached to a switching router via a high bandwidth port, such as a 1 Gbps, 10 Gbps, or a higher bandwidth port, such as high-speed Ethernet port, usually carried over a fiber, or suitable short-distance copper lines. In one embodiment, in a typical deployment using high bandwidth connections (in 2009 terms), each of about 1,000 fractional-storage servers is located on the backbone or close to the backbone and is connected to the backbone via a dedicated (guaranteed bandwidth) 1 Gbps Ethernet port, resulting in an aggregated throughput of 1,000 Gbps, which can serve about one million subscribers of standard definition streaming video, such asclient device3020, simultaneously. Such aggregated bandwidths would have required a substantially larger number of fractional-storage servers, had they been connected to other locations in the Internet, such as at edges of the Internet (close to last mile networks),Tier 3 ISPs, or at the user premises. Moreover, in some embodiments, the cost of streaming the mentioned 1,000 Gbps when the fractional-storage servers are located on the Internet backbone, or close to the Internet backbone, is expected to be significantly lower than what is expected when the servers are located elsewhere as mentioned before.
• FIG. 35 illustrates one example where an assemblingserver4020 is located at thejuncture4010 between two networks: the first network is anISP transit network4014 that connects the juncture to the Internet and provides Internet transit via aswitching router4015, and the second is alast mile network4041 that connectsend users4051 to the Internet via a switch4031 (located, for example, inside a Central Office, a Head-End, or a street-level cabinet). In one embodiment, thejuncture4010 is a network operated by a local ISP that pays transit fees for Internet traffic passing through thetransit network4014, and last mile fees for traffic passing through thelast mile network4041. A unique property of thejuncture4010 is that it is possible for an assemblingserver4020 located at the juncture to receive erasure-coded fragments sent by fractional-storage servers, such as4001 and4002, to assemble content, and to stream the content to aclient4051 via thelast mile network4041, without incurring any additional costs in comparison to other scenarios, such as where Internet packets flow from the Internet backbone to a Tier two ISP network to the Internet backbone and to the last mile network. In other words, since the assemblingserver4020 is located at the juncture, it does not create any extra traffic vianetworks4014 and4041. The assembling server can also be located at or close to an edge of the Internet, which may include the juncture, or a point aboveserver4015, such as at thetransit network4014 connecting the juncture to the Internet. When located at or close to an edge of the Internet, the assembling server has the potential not to incur additional transit fees as a result of the relaying operation, since approximately the same traffic would have to pass via the same transit network in a normal scenario. Another beneficial location for the assembling server is at the home premises, since, clearly, a relaying operation performed there does not add any significant traffic to higher levels of the network. In contrast to the above-suggested locations, in some cases an assembling server may be located at an arbitrary point on the backbone, or at other high-level points of the Internet, where it incurs additional transit fees, as fragments assembled by the server flow once over an Internet transit network going from a fractional-storage server to the assembling server, and then a second time when streamed by the assembling server to a destination client over an Internet transit network.
• In one embodiment, a distributed system is located in a few to dozens of data centers (also known as server farm or datacenter), located close to or on the Internet backbone, together housing at least 100 fractional-storage CDN servers. The servers store erasure-coded fragments associated with approximately sequential segments of streaming contents, with a storage gain of at least 5, and transmit the stored fragments on demand to assembling devices approximately according to the sequential order of the segments. In many cases, the data centers provide a convenient place to place the CDN servers close to or on the Internet backbone. A data center can be also a collocation center, or an Internet Exchange Point. In one example, a single data center can house many fractional-storage CDN servers.
• In one example, a streaming system comprising at least several hundreds of fractional-storage CDN servers located close to or on the Internet backbone, storing erasure-coded fragments encoded with a redundancy factor greater than one, and associated with approximately sequential segments of streaming contents. At least 100,000 assembling devices concurrently obtain fragments from the CDN servers, wherein the system achieves efficient load balancing and fault tolerance between the various CDN servers by determining for each of the assembling devices from which servers to obtain the fragments.
• In one example, a system comprising at least 1,000 fractional-storage CDN servers is connected to the public Internet. The servers store erasure-coded fragments associated with approximately sequential segments of streaming contents, with a storage gain greater than 5, and transmit the stored fragments on demand to assembling devices approximately according to the sequential order of the segments. Wherein the aggregated bandwidth utilized by the servers for transmitting the fragments to the assembling devices exceeds 1 Giga bit per second times the number of the CDN servers. In one optional example, the system comprises at least 10,000 fractional-storage CDN servers and the aggregated bandwidth utilized by the servers exceeds 10 Giga bit per second times the number of the CDN servers.
• In one embodiment, the assembling device categorizes the servers into two categories: (i) fastest responding servers, and (ii) slower responding servers, and approximately avoids initial fragment requests from the fastest responding servers, such that if additional fragments are needed, they are quickly retrieved from the fastest responding servers. Avoiding retrieval from the fastest responding servers when initially requesting the fragments of a segment increases the chances of retrieving a substitute fragment, needed to compensate for the lost fragments, from the fastest responding servers, and enables fast compensation that is needed for fast presentation of the streaming content. Categorizing the servers may be performed by registering measured latencies of servers responding to fragment requests by the assembling device.
• In one embodiment, a plurality of fractional-storage servers, which may be located almost anywhere around the globe, configured to store erasure-coded fragments associated with segments of streaming content. An assembling device, which may be located almost anywhere around the globe, configured to request, using a fragment pull protocol over the Internet, a set of fragments. The assembling device is further configured to compensate for lost fragments by requesting additional erasure-coded fragments that are needed to reconstruct the segments. wherein the bandwidth of the streaming content is bounded approximately only by the incoming bandwidth of the assembling device.
• In one embodiment, fractional-storage CDN servers configured to store erasure-coded fragments associated with approximately sequential segments of streaming content. An assembling device located at a point featuring an average one-way network-related latency of more than 50 milliseconds between the assembling device and the servers obtains a first set of fragments, approximately according to the sequential order of the segments, and compensates for lost fragments by obtaining a second set of erasure-coded fragments that are needed to reconstruct the segments. Wherein the bandwidth of the streaming content is bounded approximately only by the incoming bandwidth of the assembling device. Optionally, the assembling device is configured to utilize a fragment pull protocol to obtain the fragments. Optionally, the assembling device utilizes a push protocol to obtain the fragments.
• FIG. 36 illustrates one embodiment of fractional-storage servers697ato697(N), wherein each server, or a group of servers, may be owned and/or connected to the Internet in any combination and by different entities. The following 3 examples illustrate how the assembling devices balance the load on the fractional-storage servers. In oneexample servers697aand697b,server697c, and servers697(N−1) and697(N) are connected to theInternet300 via first, second, and third hosting providers (689a,689b, and689c) correspondingly. Assemblingdevice661 can select the servers regardless of the hosting provider, and in a manner that combines erasure-coded fragments from several hosting providers. At any point in time, the operator of the distributed storage system can perform a cost effectiveness analysis of the hosting and data transport services provided by each hosting provider, and look for new hosting providers as candidates for replacement of one or more of the current hosting providers. If such a replacement is found, such as when a better hosting deal can be obtained, the distributed storage operator can terminate the services of such hosting provider(s), and replace it with a better deal.
• FIG. 37 illustrates one embodiment in whichdifferent data centers1461 to1464 host fractional-storage servers. The servers store erasure-coded fragments encoded with a redundancy factor R greater than one. A plurality of assemblingdevices1499 obtain fragments needed for streaming contents. No group of servers within any one of the data centers store more than (1-1/R) of the fragments associated with a single segment to be reconstructed; meaning that if any one of the data centers stop delivering fragments, the other data centers still comprise enough erasure-coded fragments needed to decode the fragments. Upon termination of a fragment delivery service from any one of the data centers, the servers of the data center whose fragment delivery service was terminated are deselected for fragment delivery, and other servers in other data centers are selected for fragment delivery instead, while the streaming of the contents to affected assembling devices continues during this short deselection-reselection process, and without disrupting any ongoing streaming operation. Usually, each data center hosts more than one server, but a data center may also host a single server. In one embodiment, more than 200 servers are hosted in more than 20 data centers. In one embodiment, more than 10,000 servers are hosted in more than 100 data centers, and deliver fragments at an aggregated throughput of more than 10 Tera-bit per second.
• In one embodiment, the assemblingdevices1499 use a fragment pull protocol to retrieve fragments from the servers and to approach new servers, optionally on-the-fly while streaming contents, instead of servers whose data center's fragment-delivery service was terminated.
• In one embodiment, the assemblingdevices1499 use a push protocol to obtain fragments from the servers by utilizing multiple sub-transmissions. Servers whose data center's fragment-delivery service was terminated are replaced, optionally on-the-fly while streaming contents, by other servers which operate as the sub-transmission sources.
• In one example, the deselection-reselection process takes a few seconds or less. In one example, the deselection-reselection process is done faster than it takes to play the content stored in one relatively short segment. In one example, the deselection-reselection process is done by assembling devices. In one embodiment, the reselection of servers is done by a control server.
• FIG. 38 illustrates one embodiment in which fractional-storage servers withindata centers3661 to3664 store erasure-coded fragments encoded with a redundancy factor greater than one. A plurality of assemblingdevices3699 obtain decodable sets of fragments from subsets of the servers and measure fragment delivery parameters that are indicative of delivery performances, such as latency in responding to requests, or fragment loss ratios. Each assembling device can readily make the measurements on fragments sent to it. Decisions are constantly made by the assembling devices, a control server, or any other decision component, regarding selection and reselection of servers participating in the subsets. The decisions are based on the measured parameters, and are made in order to improve the measured parameters. After many such decisions are made for or by many assembling devices, it is possible to estimate the performances of the different data centers. A data center that is underperforming relative to other data centers is likely to feature one or more of the following: (i) delivers fewer fragments to assembling devices as compared to other data centers, (ii) incurs higher cost per fragment delivery, and thus is less cost effective compared to other data centers, (iii) utilizes a lower percentage of the fragment delivery bandwidth available to it, as compared to other centers, and/or (iv) exhibits any other measurable degradation in performance level, that is a result of server participation in subsets, and that can be used to differentiate it over well performing data centers. The preference of the assembling devices, or other decision component, for some servers over other servers creates a “natural selection” process that can be utilized to distinguish well performing data centers over underperforming data centers. After the data centers are distinguished, decisions can be made regarding a future utilization of each center.
• FIG. 38 andFIG. 39 illustrate some of the above principles and embodiments, in accordance with one example.Data centers3661 to3664 include fractional-storage servers.Multiple assembling devices3699 obtain erasure-coded fragments from the servers. At first, thefragment delivery throughputs3651 to3654 delivered bydata centers3661 to3664 overcommunications lines3671 to3674 respectively, are approximately equal to each other. Over time, the assembling devices measure fragment-delivery parameters associated with servers of the different centers, and select subsets of servers from which to obtain decodable sets of fragments accordingly. The measured parameters associated with center's3661 servers are not as good as the parameters measured from other servers. Center's3661 servers are therefore less frequently selected by assemblingdevices3699, and the result is a reduction in fragment delivery throughput from that center from3651 to3651′. At the same time, the measured parameters associated with center's3664 servers are better than parameters measured by assembling devices from other servers. Center's3664 servers are therefore more frequently selected by assembling devices, and the result is an increase in fragment delivery throughput from that center from3654 to3654′. The performance of the different data centers can now be compared, and decisions can be made regarding the future utilization of the center's resources. The following two examples illustrate performance comparisons and corresponding decisions.
• In the claims, a sentence such as “the erasure-coded fragments support source-selection diversity” is to be interpreted as fragments encoded using any kind of erasure-code that can produce N unique fragments, from which C combinations of decodable sets of fragments can be selected, wherein C is much greater than N. Standard parity checks, standard checksums, and standard cyclic redundancy checks (CRC) are examples of codes that do not support source-selection diversity.
• In this description, numerous specific details are set forth. However, the embodiments of the invention may be practiced without some of these specific details. In other instances, well-known hardware, software, materials, structures and techniques have not been shown in detail in order not to obscure the understanding of this description. In this description, references to “one embodiment” mean that the feature being referred to may be included in at least one embodiment of the invention. Moreover, separate references to “one embodiment” or “some embodiments” in this description do not necessarily refer to the same embodiment. Illustrated embodiments are not mutually exclusive, unless so stated and except as will be readily apparent to those of ordinary skill in the art. Thus, the invention may include any variety of combinations and/or integrations of the features of the embodiments described herein.
• Although some embodiments may depict serial operations, the embodiments may perform certain operations in parallel and/or in different orders from those depicted. Moreover, the use of repeated reference numerals and/or letters in the text and/or drawings is for the purpose of simplicity and clarity and does not in itself dictate a relationship between the various embodiments and/or configurations discussed. The embodiments are not limited in their applications to the details of the order or sequence of steps of operation of methods, or to details of implementation of devices, set in the description, drawings, or examples. Moreover, individual blocks illustrated in the figures may be functional in nature and do not necessarily correspond to discrete hardware elements. While the methods disclosed herein have been described and shown with reference to particular steps performed in a particular order, it is understood that these steps may be combined, sub-divided, or reordered to form an equivalent method without departing from the teachings of the embodiments. Accordingly, unless specifically indicated herein, the order and grouping of the steps is not a limitation of the embodiments. Furthermore, methods and mechanisms of the embodiments will sometimes be described in singular form for clarity. However, some embodiments may include multiple iterations of a method or multiple instantiations of a mechanism unless noted otherwise. For example, when a controller or an interface are disclosed in an embodiment, the scope of the embodiment is intended to also cover the use of multiple controllers or interfaces.
• Certain features of the embodiments, which may have been, for clarity, described in the context of separate embodiments, may also be provided in various combinations in a single embodiment. Conversely, various features of the embodiments, which may have been, for brevity, described in the context of a single embodiment, may also be provided separately or in any suitable sub-combination.
• Embodiments described in conjunction with specific examples are presented by way of example, and not limitation. Moreover, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art. It is to be understood that other embodiments may be utilized and structural changes may be made without departing from the scope of the embodiments. Accordingly, it is intended to embrace all such alternatives, modifications and variations that fall within the spirit and scope of the appended claims and their equivalents.

Claims (20)

1) A method for retrieving erasure-coded fragments, comprising: identifying a first group of fractional-storage servers estimated to have low response latencies in relation to an assembling device; retrieving, by the assembling device from a second group of fractional-storage servers, enough erasure-coded fragments needed for reconstructing approximately sequential segments of streaming content; while retrieving the fragments, identifying at least one server from the second group having latency higher than a certain threshold in response to a fragment pull protocol request; and using the fragment pull protocol to replace the identified server with at least one server selected from the first group.

2) The method ofclaim 1, wherein the servers are fractional-storage CDN servers, and the at least one server selected from the first group is selected approximately randomly.

3) The method ofclaim 1, wherein the servers are fractional-storage CDN servers, and the at least one server selected from the first group is selected according to a least latency criterion.

4) The method ofclaim 1, wherein the servers are fractional-storage CDN servers, and the at least one server selected from the first group is approximately the least loaded server.

5) The method ofclaim 1, further comprising while retrieving the fragments, identifying at least one server from the second group having latency variance higher than a certain threshold, and using the fragment pull protocol to replace that server with at least one server selected from the first group.

6) The method ofclaim 1, further comprising measuring the round trip delay between the assembling device and the servers using the fragment pull protocol requests.

7) The method ofclaim 1, wherein identifying the at least one server having latency higher than a certain threshold is performed by the assembling device via the fragment pull protocol, whereby no dedicated communication is needed for identifying the server.

8) The method ofclaim 1, wherein identifying the at least one server having latency higher than a certain threshold is performed by obtaining the router hop-count between the servers and the assembling device, whereby a higher hop-count usually indicates higher latency.

9) The method ofclaim 1, wherein the erasure-coding is rateless-coding potentially resulting in fragments having a limitless redundancy factor; whereby a higher redundancy enables a larger first group.

10) The method ofclaim 1, wherein identifying the servers of the first group is performed by obtaining the router hop-count between the servers and the assembling device, whereby a lower hop-count usually indicates lower latency.

11) The method ofclaim 1, wherein identifying the servers of the first group is performed by occasionally measuring the latencies between the assembling device and the servers not belonging to the second group.

12) A system comprising: a plurality of fractional-storage CDN servers connected to the public Internet; the servers store, at an average storage gain>5, erasure-coded fragments associated with approximately sequential segments of streaming contents, and are configured to respond with fragments to fragment pull protocol requests issued by assembling devices; wherein the system is configured to direct the fragment pull protocol requests of each assembling device towards servers exhibiting relatively low latencies in responding to the assembling device.

13) The system ofclaim 12, wherein the system comprises at least 100 fractional-storage CDN servers, and each assembling device is configured to select the low latency servers by itself, and servers featuring lower latency variance are usually preferred.

14) The system ofclaim 12, wherein the system comprises at least 100 fractional-storage CDN servers, and the erasure-coding is rateless-coding potentially resulting in fragments having a limitless redundancy factor.

15) The system ofclaim 12, wherein the servers are located close to or on the Internet backbone and the assembling devices are located at the user premises.

16) The system ofclaim 12, wherein the system comprises at least 100 fractional-storage CDN servers, and a central server is configured to direct the fragment pull protocol requests towards the servers exhibiting relatively low latencies and latency variance below a certain threshold.

17) A system comprising: a plurality of fractional-storage CDN servers accessed via the public Internet and spread over a wide area; the servers store, at an average storage gain>5, erasure-coded fragments associated with approximately sequential segments of streaming contents, and are configured to deliver sub-transmissions comprising fragment sequences to assembling devices; wherein there is a high variance in latency between the servers and the assembling devices due to the fact that the system is spread over a wide area, and the system is configured to match assembling devices with servers having relatively low response latencies for sub-transmissions delivery.

18) The system ofclaim 17, wherein each assembling device is configured to match itself with the servers having relatively low latencies.

19) The system ofclaim 17, wherein the system comprises at least 100 fractional-storage CDN servers, and the erasure-coding is rateless-coding potentially resulting in fragments having a limitless redundancy factor.

20) The system ofclaim 17, wherein the servers are located close to or on the Internet backbone, and the assembling devices are located at the user premises.

Images