Pod consumes more GPU resources than the given limit about gpushare-scheduler-extender HOT 4 OPEN

#51 Just saw this is the same issue

from gpushare-scheduler-extender.

Since I am using tensorflow serving for now my solution was to use the argument
"--per_process_gpu_memory_fraction"
like this

containers:
  - name: name
    image: image
    args: ["--per_process_gpu_memory_fraction=0.1"]
    resources:
       limits:
          aliyun.com/gpu-mem: 1024

from gpushare-scheduler-extender.

I dont see a problem here. the gpushare scheduler just manages allocation of GPU memory. It can not limit or enforce the actual quota/usage.

We should close this issue in favour of #93

from gpushare-scheduler-extender.

什么时候完全支持 "GPU memory isolation” ，即 显存级别的GPU隔离呢? 阿里云cGPU容器技术， 通过一个内核驱动，为容器提供了虚拟的GPU设备，从而实现了显存和算力的隔离； 这个cGPU技术能开源吗？

from gpushare-scheduler-extender.