0

I am trying to measure the system performance when spectreV1 mitigation is on. I am using stress-ng for that. So far I have tested the system with various stressors like cpu, qsort, str, matrix, tsearch, lsearch, sem, switch and mq. I am looking at bogo ops (real-time) values to compare the performance by first disabling all the mitigations in the kernel vs. only spectrev1 enabled.

I would like to know whether these stressors are enough to test the performance or do I need to add more stressor for a better comparison.

Any help would be appreciated.

awatan
  • 101
  • 4

1 Answers1

0

Micro benchmarks do not reflect the performance of real things. Test with your actual workload.

  • Load generator replaying production load on a test system, if you have such a thing.
  • Install on test servers.
  • Gradually roll out to production hosts, if you have more than one.

Once you have chosen your mitigation level and tested it some, don't delay too much in implementing it. Rely on your capacity margin to absorb the impact, and have a back out plan to undo it if necessary.

John Mahowald
  • 30,009
  • 1
  • 17
  • 32