GPU virtualization implementations generally involve one or more of the following techniques: device emulation, API remoting, fixed pass-through and mediated pass-through. Each technique presents different trade-offs regarding virtual machine to GPUconsolidation ratio,graphicsacceleration, renderingfidelity andfeature support,portability to different hardware,isolation between virtual machines, and support for suspending/resuming andlive migration.[1][4][5][6]
InAPI remoting or API forwarding, calls to graphical APIs from guest applications are forwarded to the host byremote procedure call, and the host then executes graphical commands from multiple guests using the host's GPU as a single user.[1] It may be considered a form ofparavirtualization when combined with device emulation.[7] This technique allows sharing GPU resources between multiple guests and the host when the GPU does not support hardware-assisted virtualization. It is conceptually simple to implement, but it has several disadvantages:[1]
In pure API remoting, there is little isolation between virtual machines when accessing graphical APIs; isolation can be improved using paravirtualization
Performance ranges from 86% to as low as 12% of native performance in applications that issue a large number of drawing calls perframe
A large number of APIentry points must be forwarded, and partial implementation of entry points may decrease fidelity
Applications on guest machines may be limited to few available APIs
Hypervisors usually useshared memory between guest and host to maximize performance and minimize latency. Using anetwork interface instead (a common approach indistributed rendering), third-party software can add support for specific APIs (e.g.rCUDA[8] forCUDA) or add support for typical APIs (e.g.VMGL[9] forOpenGL) when it is not supported by the hypervisor's software package, althoughnetwork delay andserializationoverhead may outweigh the benefits.
Application support from API remoting virtualization technologies
In fixed pass-through or GPU pass-through (a special case ofPCI pass-through), a GPU is accessed directly by a single virtual machine exclusively and permanently. This technique achieves 96–100% of native performance[3] and high fidelity,[1] but the acceleration provided by the GPU cannot be shared between multiple virtual machines. As such, it has the lowestconsolidation ratio and the highest cost, as each graphics-accelerated virtual machine requires an additional physical GPU.[1]
The following software technologies implement fixed pass-through:
For certain GPU models, Nvidia and AMD video card drivers attempt to detect the GPU is being accessed by a virtual machine and disable some or all GPU features.[35] NVIDIA has recently changed virtualization rules for consumer GPUs by disabling the check in GeForce Game Ready driver 465.xx and later.[36]
For NVIDIA, various architectures of desktop and laptop consumer GPUs can be passed through in various ways. For desktop graphics cards, passthrough can be done via the KVM using either the legacy or UEFI BIOS configuration via SeaBIOS and OVMF, respectively.
For desktops, most graphics cards can be passed through, although for graphics cards with the Pascal architecture or older, the VBIOS of the graphics card must be passed through in the virtual machine if the GPU is used to boot the host.[37]
For laptops, the NVIDIA driver checks for the presence of a battery via ACPI, and without a battery, an error will be returned. To avoid this, an acpitable created from text converted into Base64 is required to spoof a battery and bypass the check.[37]
For the laptop graphics cards that are Pascal and older, passthrough varies widely on the configuration of the graphics card. For laptops that do not have NVIDIA Optimus, such as the MXM variants, passthrough can be achieved through traditional methods. For laptops that have NVIDIA Optimus on as well as rendering through the CPU's integrated graphics framebuffer as opposed to its own, the passthrough is more complicated, requiring a remote rendering display or service, the use of Intel GVT-g, as well as integrating the VBIOS into the boot configuration due to the VBIOS being present in the laptop's system BIOS as opposed to the GPU itself. For laptops that have a GPU with NVIDIA Optimus and have a dedicated framebuffer, the configurations may vary. If NVIDIA Optimus can be switched off, then passthrough is possible through traditional means. However, if Optimus is the only configuration, then it is most likely that the VBIOS is present in the laptop's system BIOS, requiring the same steps as the laptop rendering only on the integrated graphics framebuffer, but an external monitor is also possible.[38]
In mediated device pass-through or full GPU virtualization, the GPU hardware providescontexts withvirtual memory ranges for each guest throughIOMMU and the hypervisor sends graphical commands from guests directly to the GPU. This technique is a form ofhardware-assisted virtualization and achieves near-native[b] performance and high fidelity. If the hardware exposes contexts as full logical devices, then guests can use any API. Otherwise, APIs and drivers must manage the additional complexity of GPU contexts. As a disadvantage, there may be little isolation between virtual machines when accessing GPU resources.[1]
The following software and hardware technologies implement mediated pass-through:
VMware Virtual Shared Pass-Through Graphics Acceleration[a] with Nvidia vGPU[42] or AMD MxGPU[43]
Citrix XenServer shared GPU with Nvidia vGPU, AMD MxGPU or Intel GVT-g[28][29]
Thincast Workstation - Virtual 3D feature (Direct X 12 & Vulkan 3D API)
While API remoting is generally available for current and older GPUs, mediated pass-through requires hardware support available only on specific devices.
Hardware support for mediated pass-through virtualization
GPU architectures are very complex and change quickly, and their internal details are often kept secret. It is generally not feasible to fully virtualize new generations of GPUs, only older and simpler generations. For example,PCem, a specialized emulator of theIBM PC architecture, can emulate aS3 ViRGE/DX graphics device, which supportsDirect3D 3, and a3dfx Voodoo2, which supportsGlide, among others.[49]
When using aVGA or anSVGA virtual display adapter,[50][51][52] the guest may not have 3D graphics acceleration, providing only minimal functionality to allow access to the machine via a graphics terminal. The emulated device may expose only basic 2D graphics modes to guests. The virtual machine manager may also provide common API implementations usingsoftware rendering to enable 3D graphics applications on the guest, albeit at speeds that may be low as 3% of hardware-accelerated native performance.[1] The following software technologies implement graphics APIs using software rendering:
^Intel GVT-g achieves 80–90% of native performance.[39][40] Nvidia vGPU achieves 88–96% of native performance considering the overhead on a VMware hypervisor.[41]
^abWalters, John; Younge, Andrew; Kang, Dong-In; Yao, Ke-Thia; Kang, Mikyung; Crago, Stephen; Fox, Geoffrey (2014). "GPU Passthrough Performance: A Comparison of KVM, Xen, VMware ESXi, and LXC for CUDA and OpenCL Applications".IEEE 7th International Conference on Cloud Computing.Anchorage:IEEE Computer Society. pp. 636–643.doi:10.1109/CLOUD.2014.90.ISBN978-1-4799-5063-8.ISSN2159-6190.
