Effect of sched_latency_ns

[Mahmood Naderan <mahmood.nt@gmail.com>]

I would like to test the effect of latency_ns parameter on the
execution time of a program. The program is a 16-thread MM
implementation and the CPU is Ryzen with 8/16 (cores/threads). I also
have created a busy system with stress --cpu 16.

The default value of latency_ns is 24,000,000. If I use lower values,
the execution time gets longer. If I use larger values, the execution
time goes a little bit down. But the problem is when I use perf and
see the stats, they vary over multiple executions. This is expected
because of real hardware execution, but I am not able to figure out
the conclusion about the effect.

Example:

For 100,000

       142,696        context-switches
         1,625        cpu-migrations
        51,106        page-faults
  40.103996890 seconds time elapsed

For 24,000,000

        62,185        context-switches
           379        cpu-migrations
        52,929        page-faults
  19.424256128 seconds time elapsed

For 100,000,000

        63,646        context-switches
           513        cpu-migrations
        52,930        page-faults
  19.716886869 seconds time elapsed

        54,644        context-switches
           487        cpu-migrations
        52,894        page-faults
  16.788766277 seconds time elapsed



Any explanation about the effect? What should I expect?



Regards,
Mahmood