Hi, I did benchmarking of scheduler fairness with enabled steal time accounting(STA) in KVM. And results are really interesting. Looks like STA provides worse scheduler fairness against disabled STA (no-steal-acc cmdline param) I created benchmark, main idea is: 2 cgroups with cpu.shares proportion like 1/9, run identical work in both groups, and expecting to get the same proportion of work done – 1/9. Condition – CPU overcommit. On bare metal it is fair +- some percentages of fluctation. On KVM with no STA it’s less fair. With STA enabled results are ugly! One again – in CPU overcommit situation. Host: ubuntu 14.04, Intel i5 – 2x2 CPUs 2 VMs (4 vCPU each) are working in parallel. 2:1 cpu overcommit. Each VM has running benchmark: cgroups cpu.shares proportion is 128/1152 (10%/90%), work – spinning in cycle, number of cycles are being counted. Results: First column – work proportion. Expected value 90% Second column – total number of loop cycles from all workers (x100000). Interval between printed lines – 1 sec VM1: STA is disabled (no-steal-acc) 85% 29683 85% 29520 87% 29574 87% 31386 84% 29578 87% 29595 86% 29644 87% 29655 87% 30059 88% 30000 90% 30037 90% 29720 90% 29951 86% 29958 88% 29971 90% 29829 82% 29841 82% 29998 82% 29752 80% 29722 88% 28825 85% 28283 85% 29746 77% 29657 82% 29674 84% 29764 87% 29567 79% 29529 84% 29633 VM2: STA is enabled 51% 30709 38% 29564 44% 29429 47% 29759 41% 29792 0% 29987 51% 52666 50% 29753 50% 29670 54% 29461 45% 29927 35% 29743 50% 29683 50% 29971 47% 30033 0% 121204 7% 116039 25% 119069 24% 40269 5% 37684 24% 30039 22% 30269 44% 31861 29% 29652 24% 29558 24% 29783 24% 29494 24% 29607 43% 29613 25% 29377 My benchmark sources: https://github.com/YustasSwamp/steal-time-benchmark Any ideas? Do you have any steal time specific benchmarks? Thanks, Alexey -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
