Re: perf: massive perf_event slowdown between 4.9 and 4.11-rc

[Peter Zijlstra]

On Thu, Mar 16, 2017 at 11:54:58AM -0400, Vince Weaver wrote:
> Hello
> 
> My student actually noticed this before I did, I was hoping it was some 
> sort of error in her data.
> 
> Anyway all perf_event functionality (especially reads) has become about 
> 20x slower, at least on Intel machines (haswell and skylake are the only 
> ones I've tested) sometime between 4.9 and 4.11-rc
> 
> For example, in the PAPI tests:
> 
> 4.11-rc2
> 
> Total cost for PAPI_read (2 counters) over 1000000 iterations
> min cycles   : 15192
> max cycles   : 3887735
> mean cycles  : 15662.057418
> std deviation: 19079.398693
> 
> 
> 4.9
> 
> Total cost for PAPI_read (2 counters) over 1000000 iterations
> min cycles   : 864
> max cycles   : 78459
> mean cycles  : 908.010315
> std deviation: 144.875697
> 
> 
> The perf_event_test validation tests are also showing this, even when 
> using rdpmc() rather than read.
> 
> Is there a likely change that might have caused this?  Hoping to avoid 
> bisecting it as that will kill the rest of the week probably.

No immediate clue; but I can have a poke. Do you have a handy way of
showing this using: perf test rdpmc, or should I hack something
together?