GPU Scheduling Workload Scenarios

GPU Idle Reclaim

In HPE Ezmeral Unified Analytics Software, consider two GPU workloads, denoted as Workload1 and Workload2. Currently, Workload1 is running and is in an idle state while Workload2 is pending due to lack of available GPU resources. In this scenario, if the idle duration of Workload1 exceeds an idle time threshold, Workload1 is preempted in favor of Workload2. Following the preemption, Workload1 goes into a pending state, while Workload2 is allocated GPU resources and starts running.

Active GPU Usage

In HPE Ezmeral Unified Analytics Software, consider two GPU workloads, denoted as Workload1 and Workload2. Currently, Workload1 is running and is using GPU resources while Workload2 is pending due to lack of available GPU resources. The custom scheduler runs a cron job every 5-10 minutes to determine the eligibility of reclaiming pods based on their GPU usage and the annotation values set in the priority class attached to the pod.

If the GPU usage for Workload1 is greater than 0.0, Workload1 cannot be preempted in favor of Workload2. In this scenario, Workload1 will continue to run and utilize the GPU resources without interruption.

If the GPU usage for Workload1 is equal to 0.0 and if the idle duration of Workload1 exceeds an idle time threshold, Workload1 is preempted in favor of Workload2. Following the preemption, Workload1 goes into a pending state, while Workload2 is allocated GPU resources and starts running.

Priority Scheduling

In HPE Ezmeral Unified Analytics Software, consider three GPU workloads, denoted as Workload1, Workload2, and Workload3. Currently, Workload1 is running and is in an idle state, Workload2 is pending due to lack of available GPU resources, and Workload3 has the highest priority among the three workloads and is pending due to lack of available GPU resources. In this scenario, if the idle duration of Workload1 exceeds an idle time threshold, Workload1 is preempted in favor of Workload3. Following the preemption, Workload1 goes into a pending state, Workload3 is allocated GPU resources and starts running, and Workload2 will continue to be in the pending state.

Notebook Example for GPU Idle Reclaim

Consider a scenario in which HPE Ezmeral Unified Analytics Software is configured with a single physical GPU. In this scenario, you have chosen the small vGPU size, which includes 7 vGPUs. Each application will always have a maximum of one vGPU assigned to it.

Now, assume you have seven notebook servers, denoted as idle-gpu-notebook, used-gpu-notebook-1, used-gpu-notebook-2, used-gpu-notebook-3, used-gpu-notebook-4, used-gpu-notebook-5, and used-gpu-notebook-6. In this scenario, the idle-notebook-gpu notebook server has an idle GPU with no GPU usage while the six other notebook servers are actively using GPU resources.

You can navigate to the GPU Control Panel screen to check the status of these notebook servers. There you can see that one notebook server has an Idle status and the six others have a Running status.

You can click Notebooks to view the details of each notebook server. You can confirm that the idle notebook has no GPU usage, and six others have an active GPU usage by clicking the GPU utilization chart icon in the Actions menu.

Consider creating another GPU-enabled notebook server, denoted as test-idle-notebook-2. As the GPU usage for idle-gpu-notebook is equal to 0.0, as soon as the idle duration of idle-gpu-notebook exceeds an idle time threshold, idle-gpu-notebook is preempted in favor of test-idle-notebook-2. Following the preemption, idle-gpu-notebook goes into a pending state, while test-idle-notebook-2 is allocated GPU resources and starts running.