On Thu, Jan 8, 2009 at 10:36 PM, Darryl Gove <darryl.g...@sun.com> wrote: > Ok. That's an entertaining problem. I'd suggest using the tools like > mpstat, prstat to look at the activity of the system during the time > when the fluctuations occur. You should also look at the dtrace toolkit > since that covers a wide range of scenarios. > http://opensolaris.org/os/community/dtrace/dtracetoolkit/
In particular, "prstat -mL" is likely to give better detail. When -m is NOT used, the cpu utilization displayed is a time decayed average. That may be useful in some cases, but is completely worthless when you are trying to observe short-lived events. The -L option will make it so the data is presented on a per LWP (thread) basis. Note that the CPU times displayed will be relative to a single CPU strand. That is, when "prstat" shows that a process is using 100% of the CPU, that would mean that it is using 100% of all CPU's. When "prstat -mL" shows that a LWP is using 100%, it is using 100% of a single CPU (or core, or hardware strand). In other words, "prstat -mL" on an 8 core T2000 can show 32 threads running at 100%. -- Mike Gerdts http://mgerdts.blogspot.com/ _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
