On Wed, Apr 11, 2012 at 02:18:14PM -0400, Richard Lowe wrote: > > Different application algorithms show different high-runners but the > > high-runner locks are usually > > not called very often but are held for an abnormally long time. For some > > algorithms the high-runners are in libc (e.g. malloc) and OpenMP but in some > > others it is my application's explicit locks. > > > > Based on something Hans mentioned, something approaching: > > cputrack -c PAPI_l1_icm,PAPI_l2_ich,PAPI_l2_icm <your application> > > Could be useful in comparison to a system on which this doesn't occur > (a synthetic test case would be better, I'm not sure these counters > would be so useful under the real workload). > > There's also the DTrace cpc provider, but I can't currently come up > with a good way to make use of it for what we'd want to know.
FYI, there is a white paper about the L1I cache aliasing issue on family 0x15 here: http://developer.amd.com/Assets/SharedL1InstructionCacheonAMD15hCPU.pdf I don't know whether this applies to Illumos or this particular problem at all, but it was the first thing I remembered when I read "slow" and "Opteron" :) There was also a lock performance problem which also affected older AMD CPU families, but it was specific to the mutex implementation in glibc. The mutexes were spinning for a too short time, increasing that time would massively reduce the contention seen. Although our default spin count for adaptive mutexes seems to be 10 times what glibc used, using a higher value as described in http://src.illumos.org/source/xref/illumos-gate/usr/src/lib/libc/port/threads/synch.c#88 could be worth a try. Hans -- %SYSTEM-F-ANARCHISM, The operating system has been overthrown ------------------------------------------- illumos-discuss Archives: https://www.listbox.com/member/archive/182180/=now RSS Feed: https://www.listbox.com/member/archive/rss/182180/21175430-2e6923be Modify Your Subscription: https://www.listbox.com/member/?member_id=21175430&id_secret=21175430-6a77cda4 Powered by Listbox: http://www.listbox.com
