Hi Rodrigo, You are right, in that it very much depends on the nature of the workload, the system configuration, and a variety of other factors. But over time I (and others i'm sure) have tended to correlate "typical" values that a tool like mpstat provides with certain workload / system load characteristics.
Every now and than I happen across some data that just doesn't jive with my expectations, and that may be because: - My understanding of the workload is (or has become) incorrect - Something about the system has changed, for better or worse. At that point, tools like DTrace(1m) can be used to dive deeper to understand why mpstat is saying what it is. Sometimes we learn something new about the workload, sometimes we find a bug or suboptimal part of the system, and sometimes we realize that our expectations were just off. :) It's hard for me to say whether 773 migrations/sec is indicative of a condition where performance is being left on the table, without knowing more about the workload, the system, etc. Thread migrations in and of themselves are a perfectly normal part of the system as it load balances and multiplexes many threads across (usually many fewer) CPUs. But you want to look for situations where threads are "needlessly" migrating....as this will lower your overall cache utilization while not necessarily improving overall CPU utilization. Differentiating good vs needless migrations using only a high level tool like mpstat is difficult...DTrace's sched provider can be helpful here. Looking at your data, one avenue I might suggest pursuing is understanding where those involuntary context switches (preemptions), are coming from. I notice that the icsw rate and migr rate are strikingly similar, so they could be related. When a thread is preempted, it goes through the dispatcher and may be enqueued to run on a different CPU....or it could be enqueued on the same CPU, but other idle CPUs can steal the thread...either way, that's a migration. -Eric This message posted from opensolaris.org _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
