This document gives an overview of future reservation requests in calendar mode.

Use future reservation requests in calendar mode to obtain high-demandresources, such as for creating virtual machine (VM) instances that have GPUs orTPUs attached. When Google Cloud approves a reservation request,Compute Engine provisions your reserved resources at your specified date andtime, and for a duration of up to 90 days. You can then use the reservedresources for creating GPU VMs, H4D VMs, or TPU VMs to run the followingworkloads:

• Model pre-training jobs

• Model fine-tuning jobs

• High performance computing (HPC) simulation workloads

• Short-term expected increases in inference workloads

For more information about other ways to reserve resources inCompute Engine, seeChoose a reservation type.

Create a request in calendar mode

The following sections explain how to view resource availability, as well aswhat details to specify when you create a future reservation request in calendarmode.

View resource future availability

Before you create a future reservation request in calendar mode, you can viewthe future availability in a region of the following resources:

• For GPU or H4D VMs, up to 60 days in advance

• For TPUs, up to 120 days in advance

Compute Engine uses theDynamic Workload Scheduler (DWS)to view when your requested resources are available. When you create a request,specify the number, type, and reservation period for the resources that youconfirmed as available. Google Cloud is more likely to approve yourrequest if you supply this information.

Define request properties

When you create a future reservation request in calendar mode, you must specifythe following properties:

• Auto-delete. This property determines if Compute Engine deletesthe automatically created (auto-created) reservation for your request atthe end time, even if the reservation isn't fully consumed. To create arequest in calendar mode, you must enable the auto-delete option.

• Consumption type. This property defines how VMs consume the auto-createdreservation. When you create a request in calendar mode, you must specifythat you want to createspecifically-targeted reservations.This setting means that only VMs that target the reservation can consume it.

• Deployment type. This property defines the collocation of your reservedresources. Compute Engine reserves resources based on the resourcetype:

  • For GPU or H4D VMs, you must specify the dense (DENSE) deploymenttype when you create a request. This configuration specifies to denselyreserve resources for minimal network latency.

    Important: Compute Engine densely allocates resources onlywithin the auto-created reservation for a single request. Auto-createdreservations from multiple requests might not be located near eachother.

  • For TPUs, Compute Engine uses the flexible (FLEXIBLE)deployment type by default. This configuration specifies to reserveresources as close as possible on a best-effort basis.

• Name. The name of your request, which must be unique within yourproject.

• Number of resources. The number of GPU or H4D VMs or TPUs to reserve atyour requested start time.

• Planning status. This property defines if you immediately submit yourrequest to Google Cloud for review, or save it as a draft and submitit later. When you create a request in calendar mode, you must specify toimmediately submit the request for review.

• Reservation mode. This property defines the method to reserve resources,which you must set toCALENDAR for a request in calendar mode.

• Reservation name. The name for the reservation thatCompute Engine automatically creates if Google Cloud approvesyour request.

• Share type. This property defines if other projects in your organizationcan consume the auto-created reservation for your approved request. You canspecify one of the following options:

  • Single-project. Only your project can consume the reserved capacity.

  • Shared. You can share the reserved capacity with up to 100 otherprojects in your organization. If you specify this option, then you mustspecify the projects to share the auto-created reservation with. Formore information, see thebest practices for shared reservations.

• Reservation period. The date and time when Compute Engineprovisions your requested capacity, and you can consume it. The reservationperiod includes the following:

  • Start time. When you want to start consuming your reserved capacity.Based on the resources that you reserve, the start time must be at leastone of the following values from when you create and submit a request:

    • For GPU and H4D VMs, 87 hours (three days and 15 hours)

    • For TPUs, six hours

  • End time. When your requested capacity is no longer reserved foryou. At this time, Compute Engine deletes the auto-createdreservation, and stops or deletes any VMs that consume the reservationbased on thetermination actionthat you specified for the VMs.

• Resource properties. The hardware requirements of the GPU VMs, H4D VMs,or TPUs that you want to reserve. VMs can only use a reservation if theirproperties match the reservation's properties. For more information, see therequirements to consume reservations.

• Workload type. If you reserveTPU v5e, then you mustspecify how to reserve capacity based on your workload type:

  • Batch. For workloads that handle large amounts of data in single ormultiple operations, such as machine learning (ML) training workloads.

  • Serving. For workloads that handle concurrent requests and requireminimal network latency, such as ML inference workloads.

• Zone. The zone where you want to reserve capacity.

Request review process

To reserve capacity by using a future reservation request in calendar mode, youmust create and submit the request to Google Cloud for review. After youcreate and submit a request, Google Cloud reviews it within a minute, andthen one of the following occurs:

• Google Cloud approves your request: Compute Enginereserves your requested resources and, within a minute after approval,automatically creates an empty reservation. At the request start time,Compute Engine provisions your requested capacity by increasing thenumber of GPU VMs, H4D VMs, or TPUs in the reservation.

  Caution: After you create a request, you can't cancel, delete, or modify it.You commit to pay for the requested capacity at the request start time,regardless if you use the capacity or not.

• You encounter an error. The request fails because the request's zonelacks sufficient resources. We recommend that you view future resourceavailability again, and then create and submit a new request for review.

Request lifecycle

The following diagram shows the different states that Compute Engine canset a future reservation request in calendar mode to:

The states and flow of events shown in the preceding diagram are as follows:

• PENDING_APPROVAL: you created and submitted a request for review.Within a minute, Google Cloud approves the request.

• APPROVED: Google Cloud approved your request. Then, within aminute, Compute Engine automatically creates an empty reservationand changes the request state toPROCURING.

• PROCURING: Compute Engine schedules the provisioning of yourreserved resources. Before the request start time, the request state changestoPROVISIONING.

• PROVISIONING: Compute Engine is provisioning your reservedresources by increasing the number of reserved GPU VMs, H4D VMs, or TPUs inthe auto-created reservation. At the request start time, the request statechanges toFULFILLED.

• FULFILLED: Compute Engine has provisioned your reservedresources, and you're charged for them. You can consume the auto-createdreservation by creating VMs until the request end time.

At the request end time, Compute Engine deletes the request and theauto-created reservation. It also stops or deletes any VMs that consume thereservation based on thetermination actionthat you specified for the VMs.

Consume provisioned capacity

After Google Cloud approves a future reservation request in calendar mode,Compute Engine automatically creates a reservation with the followingcharacteristics:

• The auto-created reservation has no reserved GPU VMs, H4D VMs, or TPUs; youcan't consume it yet.

• The auto-created reservation inherits the VM or TPU properties specified inyour request.

At the request start time, Compute Engine provisions your requestedcapacity by increasing the number of GPU VMs, H4D VMs, or TPUs in theauto-created reservation. You can then consume the reservation by creating GPUVMs, H4D VMs, or TPU VMs that meet all of the following conditions:

• The VMs and the reservation havematching properties.

• The VMsspecifically target the reservation.

• The VMs use thereservation-bound provisioning model.

• The VMs must bestopped or deleted at the reservation end time.

You can create VMs until the reservation is fully consumed or until the requestend time. At the request end time, Compute Engine deletes theauto-created reservation, and stops or deletes any VMs that consume thereservation.

Quota

Future reservation requests in calendar mode must use thereservation-bound provisioning model.This model doesn't require Compute Engine quota to reserve resources.However, before you create a request, verify that you have sufficient quota forany resources that aren't part of a reservation when you create VMs, such asdisks or IP addresses.

Pricing

When you create and submit a future reservation request in calendar mode, andGoogle Cloud approves your request, you don't immediately incur charges.Instead, you incur charges when the following occurs:

• Compute Engine provisions your requested capacity. When yourrequest reaches theFULFILLED state at the request's start time, you incurcharges for the provisioned resources according toDWS pricing. This pricingmodel offers vCPUs, memory, GPUs, and TPUs at a discounted price compared tostandard pricing.

• You use resources outside of the reservation. When you create VMs thatconsume an auto-created reservation, you don't incur additional charges forthe consumed resources. You only incur charges for resources that aren'tpart of the reservation, such as disks or IP addresses.

You stop incurring charges for the reserved resources at the request end time.At this time, Compute Engine deletes the auto-created reservation, andstops or deletes any VMs that consume the reservation based on their terminationaction.

• You can reserve the following number of VMs per request:

  • For GPU VMs, between 1 and 80 VMs

  • For H4D VMs, between 1 and 256 VMs

• You can reserve the following machine series:

  • A4

  • A3 Ultra

  • A3 Mega

  • A3 High with 8 GPUs

  • H4D

• You can reserve GPU VMs only inspecific zones. For H4Dregional availability, seeAvailable regions and zones and usetheMachine series filter to view only the zones where you can reserveH4D instances.

• You can't use an instance template to create requests for GPU or H4D VMs.

Limitations for requests for TPUs

You can only reserve TPUs as follows:

• You can reserve 1, 4, 8, 16, 32, 64, 128, 256, 512, or 1,024 TPU chips perrequest.

• You can reserve the following TPU versions:

  • TPU7x (Preview)

  • TPU v6e

  • TPU v5p

  • TPU v5e

• You can only reserve 1, 4, or 8 TPU v5e chips for serving (SERVING)workload types.

• You can reserve TPUs only in the following zones:

  • TPU7x:

    • us-central1-c

  • TPU v6e:

    • asia-northeast1-b

    • europe-west4-a

    • us-east5-a

    • us-east5-b

    • us-south1-ai1b

  • TPU v5p:

    • us-east5-a

  • TPU v5e:

    • For batch (BATCH) workload types:

      • europe-west4-b

      • us-west4-b

    • For serving (SERVING) workload types:

      • us-south1-a

Limitations for all auto-created reservations

An auto-created reservation for a request has the following limitations:

• You can only modify the reservation as follows:

  • To allow or disallow Vertex AI jobs from consuming it.

  • After the reservation start time.

• You can't apply committed use discounts (CUDs) or sustained use discounts(SUDs) to the reservation.

• You can't delete the reservation; Compute Engine deletes it at theend time for the reservation.

What's next

• Create a future reservation request in calendar mode