Skip to content

Latest commit

 

History

History
144 lines (81 loc) · 7.78 KB

vGPU.md

File metadata and controls

144 lines (81 loc) · 7.78 KB
Raw

Virtual GPU Support

Graphistry is designed to run on virtual GPUs using Nvidia's vGPU capabilities. This may simplify administration, encourage additional privacy modes, and improve isolation. It is tested on Nutanix, Vmware, and whatever is used by the major cloud providers.

If you are installing on such a system, we are happy to help, so please reach out.

In general, running Graphistry with a hypervisor requires either no special configuration, or with a vGPU, a 1-line setting in data/config/custom.env. However, virtual environment installation may take considerably more planning and effort when unfamiliar. In particular, for steps around virtualization planning, passthrough/vGPU provisioning, vGPU licensing, vGPU driver installation, and surrounding testing and troubleshooting.

Architectural considerations: You may not want vGPUs

We recommend checking whether you want baremetal OS, hypervisor with a passthrough GPU driver, or hypervisor with a vGPU driver.

A baremetal OS (no hypervisor) or passthrough driver (hypervisor with non-vGPU driver) configuration may be sufficient, meaning you can skip the complexity of using vGPU drivers and licensing:

  • Hypervisors: If only 1 dedicated GPU application, or multiple applications running in the same guest OS, you can expose it in passthrough mode
  • Scaling users through multiple GPUs:
    • Graphistry already automatically uses all GPUs exposed to it, primarily for scaling to more user sessions
    • New APIs are starting to use multi-GPUs for acceleration as well
  • Multiple Graphistry installs
    • You can launch concurrent instances of Graphistry using docker: docker-compose up -p my_unique_namespace_123
    • You can configure docker to use different GPUs or share the same ones
  • Isolate Graphistry from other GPU software
    • Docker allows picking which GPUs + CPUs are used
    • ... For both sharing and isolation
  • Nvidia vGPUs do not yet support Uniform CPU/GPU memory sharing, which causes RAPIDS processes to crash on bigger-than-GPU memory use instead of transparently spilling

Longer-term, Graphistry is aiming to push most/all GPU use to Dask, which adds even more flexibility for resource sharing.

GPU Virtualization configuration planning

Plan licensing and driver versons first, as mistakes may require starting over.

Most likely, you'll want vGPU profile 8Q with vGPU 11.0+ (11.5+ for AI) vGPU drivers:

GPU Driver

You will install a hypervisor GPU driver in the hypervisor and a guest OS GPU driver in the guest OS:

Virtualization features: C (vCS)

Nvidia vGPUs have different labels (A, B, C, Q, ...) that correspond with enabled features.

Only C (vCS) and Q (Quadro vDWS) support OpenCL/CUDA. The compute profile (C) is intended for compute workloads like Graphistry. See official NVIDIA CUDA Toolkit and OpenCL Support on NVIDIA vGPU Software.

Virtualization type: Logical (not MIG)

Only one kind of vGPU virtualization currently works with Nvidia RAPIDS:

  1. Supported - Logical time and memory sliced: Nvidia RAPIDS-compatible vGPUs are for 11.0 (11.5+ for AI) . Each vGPU gets time-sliced and a maximum amount of memory.

  2. Unsupported - MIG (physical virtualization): vGPU 11.1+ supports more true isolation of GPU cores and memory... but Nvidia RAPIDS does not yet officially work on MIGs.

Size: 2 or 8

Each GPU can be split into homogenous sizes. Ex: 1 x 8Q = 8Q and 4 x 2Q = 8Q. You may be able to oversubscribe, e.g., 4 x 4Q = 16Q.

In the case of multi-GPU systems, you can use different partition sizes on different GPUs. Ex: (1 x 8Q) + (4 x 2Q).

vGPU Software setup

Supporting vGPUs means setting up a Nvidia license manager server, GPU-capable hypervisor, a GPU-capable guest OS, and making them work together. Almost all other environment configuration is the same as regular self-hosted Graphistry setup.

There is also a 1 line change for making Graphistry work with vGPUs.

Licensing server and license

Setup an Nvidia licensing server and use it to generate and download a license. It must stay on while using your vGPU as the license gets dynamically checked.

Unlicensed GPUs will appear under nvidia-smi and be marked as unlicensed in nvidia-smi -q, but fail upon use with CUDA.

If you have a license server but cannot use it to generate a license, you may be able to receive a time/account-limited evaluation license at nvidia.com.

Hyperviser

Install hypervisor GPU drivers from the vGPU software version

Guest OS

Install guest OS GPU drivers from the vGPU software version

Configure a local license file at /etc/nvidia/gridd.conf. Take care to specify vGPU profile vCS (C) / Quadro vDWS (Q) and the right license manager server. Restart the local daemon via sudo service nvidia-gridd restart.

To confirm license activation, check the license status in nvidia-smi -q and debug logs (sudo grep gridd /var/log/messages)

Graphistry

Set RMM_ALLOCATOR=default in your data/config/custom.env to avoid relying on CUDA Unified Memory for handling bigger-than-memory workloads, which Nvidia vGPUs do not currently support.

Hyperviser-specific options

See official docs and support forums for guidance specific to your hypervisor.

Nutanix

Follow the standard Graphistry + vGPU instructions

Helpful links and configurations:

Testing

Script

In your configured VM, after loading the Graphistry containers, go to your Graphistry release folder and run ./test-gpu.sh. For more information on what this does, see quick testing and test-gpu.

Hypevisor

  • nvidia-smi should report a GPU of the expected driver version

Guest OS

  • GPU: nvidia-smi should report a GPU of the expected driver version
  • Nvidia license server approval: sudo grep gridd /var/log/messages
  • See above test-gpu.sh

Docker

  • See above test-gpu.sh

Common errors

  • No GPU driver in hypervisor / guest VM
  • Mismatching GPU driver between hypervisor / guest VM; use a supplied pair
  • RAPIDS-incompatible GPU driver version
  • Unlicensed GPU: No license, license server down, ...
  • Inappropriate license: CUDA-incapable vGPU profile/license; need C, Q

In addition, the usual GPU setup surprises may apply:

  • No Docker GPU runtime
  • Docker GPU runtime not set as default (docker info and /etc/docker/daemon.json)
  • See above test-gpu.sh discussions