Kubeshare prototyping and compute sharing deep dive

[Fizzbb]

No description provided.

[pint1022]

Comparing with GaiaGPU solution, I think Kubeshare's architecture has some good designs to monitor/manage the compute resources.

[Fizzbb]

To do: Run a pod with 50% utilization request with KubeShare, and monitor GPU utilization fluctuation.

[pint1022]

Gemini algorithm major features:

1. vgpu hook is assigned by container. it is event-driven; communicates with the gpu-isolation-module by ipc. major job: kernel watchdog, heartbeat, interrupt overuse.
2. gpu-isolation-module is a daemon process. it is event-driven; handles ipc requests; creates new sharepod; issue monitor threads by pods. major algorithm: usage-monitoring; sliding-window; request-queue; priority-management.

[pint1022]

Gemini algorithm is deserved to be further explored and tested. I like it very much. The next step analysis involves with a lot of coding and testing. two main questions: 1 Will GPU be idle if one pod runs out of quota and there is no other pod/thread asking for GPU? answer: normally it doesn't run into situation. even it runs into that unique situation, it should be avoidable by changing parameters. 2. does cuda do context switching when the pod is changed? It is vGPU driver's question, not Kubeshare's question. In my research, vGPU driver should use one context among all sharing threads.