Description of GPU Exclusive and Shared Instances
Last Updated：2022-01-14
Suppose the cluster supports the sharing and isolation of GPU computing power and graphic memory. In that case, you can decide whether to exclusively occupy or share GPU resources based on the submitted YMAL when you create a task.
Resource Description
| Resource Name | Type | Units | Description |
|---|---|---|---|
| baidu.com/v100_32g_cgpu | int64 | 1 | Number of GPU cards. Enter 1 in sharing scenarios. |
| baidu.com/v100_32g_cgpu_core | int64 | 1% | GPU card computing power, e.g.，100=Total computing power of a single card 10=One-tenth of single-card computing power |
| baidu.com/v100_32g_cgpu_memory | int64 | GiB | Graphic memory of the GPU card |
Resource Application
Single-card Exclusive Instance
resources:
requests:
baidu.com/v100_32g_cgpu: 1 // 1 card
cpu: "4"
memory: 60Gi
limits:
baidu.com/v100_32g_cgpu: 1 // limit must be consistent with the request.
cpu: "4"
memory: 60GiMultiple-card Exclusive Instance:
resources:
requests:
baidu.com/v100_32g_cgpu: 2 // 2 card
cpu: "4"
memory: 60Gi
limits:
baidu.com/v100_32g_cgpu: 2 // limit must be consistent with the request.
cpu: "4"
memory: 60GiSingle-card Sharing (No Computing Power Isolation But Graphic Memory Isolation) Instance
resources:
requests:
baidu.com/v100_32g_cgpu: 1 // 1 card
baidu.com/v100_32g_cgpu_memory: 10 // 10GB
cpu: "4"
memory: 60Gi
limits:
baidu.com/v100_32g_cgpu: 1 // limit must be consistent with the request.
baidu.com/v100_32g_cgpu_memory: 10
cpu: "4"
memory: 60GiSingle-card Sharing [Support Both Graphic Memory Isolation and Computing Power Isolation] Example:
resources:
requests:
baidu.com/v100_32g_cgpu: 1 // 1 card
baidu.com/v100_32g_cgpu_core: 50 // 50%, 0.5 card computing power
baidu.com/v100_32g_cgpu_memory: 10 // 10GB
cpu: "4"
memory: 60Gi
limits:
baidu.com/v100_32g_cgpu: 1 // limit must be consistent with the request.
baidu.com/v100_32g_cgpu_core: 50 //
baidu.com/v100_32g_cgpu_memory: 10
cpu: "4"
memory: 60Gi