gpu-memory is just for a gpu memery claim?

[hougangliu]

for now, “volcano.sh/gpu-memory” can essentially only set 3 environment variables based on https://github.com/volcano-sh/devices/blob/master/pkg/plugin/nvidia/server.go#L324 to tell user how much gpu memory he can use in current container?

for example, when set “volcano.sh/gpu-memory: 1024”, environment variables looks like below in container, in fact, process in this container can use at most 4096M gpu memory, even VOLCANO_GPU_ALLOCATED is 1024?

NVIDIA_VISIBLE_DEVICES: “0” 
VOLCANO_GPU_ALLOCATED: “1024” 
VOLCANO_GPU_TOTAL: “4096”

can anyone help me confirm if I understand correctly?

[Thor-wl]

/cc @wpeng102

[wpeng102]

Currently it only support schedule pod base on pod's “volcano.sh/gpu-memory”. Not support memory hard limit and isolation.