GCS: periodically update cgroup limits

[ryankeithster]

Summary

Currently, pod cgroup memory limits are set once during guest boot, and then never changed, not even in the case of a memory resize event. This can result is containers being OOM killed even when they have not come close to exceeding the available memory. These changes add a task that runs every 60 seconds in the GCS that polls for memory changes and then adjusts the pod cgroup memory limit accordingly.

Changes

Add periodic cgroup resize task to (cmd/gcs/main.go)

• Task runs every 60 seconds and adjusts the pod cgroup memory limit if the gcs host system total memory has changed.

New test coverage (cmd/gcs/main_test.go)

[rawahars]

This can have a couple of side-effects-

• Say an update request updated the pod which exceeded the available limit. We might return successfully but later this new code path will fail.
• There is a maximum gap of 60 seconds which is error prone.

Instead of this, can we simply perform the update of pod cgroup as part of the actual Update request?

Also, perhaps the Update request targeting pod-id should be explicitly used by caller to update the pod limit. I feel that GCS should not perform any implicit operation.