Re: Skylake minimal counts

[William C. <wc...@re...>]

On 12/04/2017 11:19 AM, Michael Petlan wrote:
> Hi all!
> 
> During testing on a Skylake-SP, I have noticed that most of the Skylake
> events have "minimum:2000003" set, even for events that used to have much
> lower minimum counts on various older micro-architectures, such as Haswell,
> Ivy Bridge, etc.
> 
> An example might be cpu_clk_unhalted which has usually 6000, except of the
> following models:
> 
> Goldmont:     minimum:2000003
> Skylake:      minimum:2000003
> Westmere:     minimum:100000
> 
> Similarly, for inst_retired, for previous CPUs, it was usually 6000, except of:
> 
> Broadwell:    minimum:2000003
> Goldmont:     minimum:2000003
> Haswell:      minimum:2000003
> Sandy Bridge: minimum:2000000
> Skylake:      minimum:2000003
> Westmere:     minimum:2000003
> 
> (e.g. Atom, Core2, Nehalem are 6000)


Hi Michael,

Setting the clk_unhalted minimum to 6000 would be unwise.  The interrupt handler to collect a sample is on the order of a few thousand cycles.  The clk_unhalted and inst_retired should probably stay in the range they are in, maybe be a bit lower.  For infrequent events such as baclear limits in the millions would too high and require someone to wait a long time to get enough samples.  The general idea is want lower limit to limit the number of samples to around a thousand samples a second on a processor.

-Will


> 
> I know that the minimal counts are nothing precise and were determined rather
> "randomly", but for Skylakes, it seems that 2000003 is just the "default for
> everything" except of some exceptions...
> 
> Since OProfile does not allow users to enforce a lower value there, the 2000003
> values seem to be pretty limiting, don't they?
> 
> Shouldn't they be re-configured? If you think so, I might do that. However,
> in such case, I am curious how would you do it (which load/criteria would you
> use)...
> 
> Cheers,
> Michael
>