@pstats frequently reports 0 cpu cycles

[topolarity]

julia> @pstats rand(1000,1000)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
┌ cpu-cycles               0.00e+00   47.4%  #  0.0 cycles per ns
│ stalled-cycles-frontend  0.00e+00   47.4%  #  NaN% of cycles
└ stalled-cycles-backend   0.00e+00   47.4%  #  NaN% of cycles
┌ instructions             1.50e+07   52.7%  #  Inf insns per cycle
│ branch-instructions      2.38e+05   52.7%  #  1.6% of insns
└ branch-misses            8.35e+01   52.7%  #  0.0% of branch insns
┌ task-clock               3.81e+06  100.0%  #  3.8 ms
│ context-switches         0.00e+00  100.0%
│ cpu-migrations           0.00e+00  100.0%
└ page-faults              0.00e+00  100.0%
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

This happens for me about ~20-50% of the time.

I'm not sure why it thinks it measured 47% of the running time but measured no cycles at all.

[Zentrik]

Maybe all your PMU counters are getting used up?

[topolarity]

That seems likely, but I still thought the effective running time calculation should have given us an indication (i.e. 47.4% should have been 0.0% and there should have been a warning)

[topolarity]

@Zentrik I'm worried this might be related to the usage of PR_TASK_PERF_EVENTS_DISABLE

If I apply this diff:

diff --git a/src/LinuxPerf.jl b/src/LinuxPerf.jl
index 7326cb3..a5b1742 100644
--- a/src/LinuxPerf.jl
+++ b/src/LinuxPerf.jl
@@ -1138,9 +1138,9 @@ macro pstats(args...)
             @debug dump_groups(groups)
             bench = make_bench_threaded(groups, threads = $(opts.threads))
             try
-                enable_all!()
+                enable!(bench)
                 val = $(esc(expr))
-                disable_all!()
+                disable!(bench)
                 # trick the compiler not to eliminate the code
                 @static if isdefined(Base, :donotdelete)
                     Base.donotdelete(val)

then the problem appears to go away.

Maybe the global disable is not accounting for being in the middle of a measurement, etc. so the timers become inaccurate?

[topolarity]

It's possible this is a kernel bug, so it'd be good to check on more HW / kernel versions to see if this is fixed on newer kernels or different hardware.