tracelog vs. xperf

[mikedn]

The xperf that ships with the current Windows SDK doesn't seem to have the -pmcprofile option so it's not possible to use it to configure the profile source. Options:

• Remove the option. It looks like InstructionsRetired is the default in my experiments
• Replace xperf with tracelog:

tracelog -start pmc_counters -eflag PROC_THREAD+LOADER+PMC_PROFILE+PROFILE -ProfileSource InstructionRetired -f pmc.etl

That said, it's not clear to me how useful this PMC_PROFILE thing is. As far as I can tell this alone can't give accurate numbers due to thread switching. Or perhaps I'm missing something?

[mikedn]

cc @AndyAyersMS

[AndyAyersMS]

Thanks for pointing out a non-xperf way of running this.

I've found the Instructions retired data to be quite useful, especially for jit TP analysis. I have gotten very stable numbers this way (~ 0.1% variation from run to run in most cases).

[mikedn]

I've found the Instructions retired data to be quite useful, especially for jit TP analysis. I have gotten very stable numbers this way (~ 0.1% variation from run to run in most cases).

Indeed, it works very well. The sampling frequency is so high that errors resulting from (usually infrequent) thread switching are likely minimal.

I also tried to use the PMC information reported via CSWITCH events and the results are practically identical - ~0.02% more instructions, possibly due to kernel work (ISR/APC/DPC stuff). However, CSWITCH events occasionally fail to work correctly, they report significantly fewer instructions. Oddly enough, this only happens when no other applications are running on the machine. PMC_PROFILE appears to work correctly in all circumstances.