Skip to content

Rebalancing

Andreas Ellisonrequested to merge
rebalancing-metric into main

What

This PR adds two variations of CFS that support rebalancing threads among CPUs in a container.

  • CFSSimpleStealing: Whenever a CPU becomes idle, it steals (about) half of the run queue of another CPU, chosen at random, by weight.
  • CFSLatencyRebalancing: Every CPU tries to balance out "load" every REBALANCING_TICK nanoseconds by stealing threads from the CPU with largest load, if necessary. The design for this is documented in the Rebalancing design doc.

Testing

Currently, there is only one test for the CFSSimpleStealing scheduler, which makes sure that the stealing works to some extent. I wasn't able to come up with other reasonable tests to make sure the schedulers behave correctly.

To really test them and, importantly, evaluate how well our design for latency-based rebalancing works, we would need to run some benchmarks, and we currently don't support enough syscalls to run any. The design doc contains some suggestions for benchmarks to use.

Edited by Andreas Ellison

Merge request reports