I did a reduced testcase and assuming that the reduced testcase has the same issue, it turns out that a lower sched_migration_cost_ns does not solve a specific problem, instead it seems to make a different problem less problematic. In the end the problem seemed to be worse on KVM hosts (guest changes did also help but much less so). In the end what did help was to improve the behaviour of yield_to from KVM. See the patch for more details. The problem seems to be real, my solution might not be the best one - I am open for better ways to code things.
