Skip to content

GPU Quota Management

This section describes how to use vGPU capabilities on the DCE platform.

Prerequisites

The corresponding GPU driver (NVIDIA GPU, NVIDIA MIG, Days, Ascend) has been deployed on the current cluster either through an Operator or manually.

Procedure

Follow these steps to manage GPU quotas in DCE:

  1. Go to Namespaces and click Quota Management to configure the GPU resources that can be used by a specific namespace.

  2. The GPU types supported by namespace quota management are: NVIDIA vGPU, NVIDIA MIG, Iluvatar, and Huawei Ascend.

  3. NVIDIA vGPU Quota Management: Configure the specific quota that can be used. This will create a ResourceQuota CR:

    • GPU Computing Power (nvidia.com/gpucores): The total GPU computing power that the namespace can allocate. It is the sum across multiple GPUs and can therefore exceed 100%.
    • GPU Memory (nvidia.com/gpumem): The total GPU memory that the namespace can allocate. It is the sum of GPU memory across multiple GPUs.

    Configure namespace GPU quotas

    Note

    A separate physical GPU count is no longer required in the new quota model.

  4. A namespace GPU quota is allocated from the workspace shared-resource quota:

    • Total Quotas: The GPU quota set for the current namespace.
    • Allocatable Quota: The GPU quota currently available to the namespace from the workspace shared resources.
    • The sum of all namespace quotas cannot exceed the workspace shared-resource quota. Otherwise, the namespace cannot be bound to the workspace.

Comments