Memory Tiering: cold start of workloads.

[kad]

Idea to try to implement and measure: For workloads that have both enabled DRAM+PMEM, we can try to start containers with only PMEM type of memory enabled, and enable DRAM memory nodes access at the point when container is in healthy running state.

Triggers that would allow DRAM usage

• [x] Timer: the timeout can be passed as annotation. Implemented in #293
• [ ] CRI-RM can watch for ListContainerStats messages and in case it detects that memory->working_set_bytes is not growing or growing at small rate, trigger enablement of the DRAM
• [ ] Combination of above - trigger usage of DRAM based on memory stats, but not longer than specified timeout.

[kad]

@ipuustin we have in ListContainerStats calls, which have memory->working_set_bytes field. We can utilize that field to trigger event on "now it is good time to enable DRAM".

[msivosuo]

@kad - can we close?

[msivosuo]

@kad -please refine description to include the parts not yet implemented.