On Apr 10, 2008, at 11:41 AM, Zeljko Vrba wrote: > On Thu, Apr 10, 2008 at 10:48:06AM -0400, Jonathan Adams wrote: >> >> >> Do you have administrator access to this system? With dtrace(1M), >> you >> can drill down on this, and get more data on what's going on. See: >> > Yes, I do; backtraces are rather long, so I won't post them, unless > you > explicitly want. In summary, for 8192 threads: > > - 433 hits in unix`idle > - 179 hits in libc.so`cond_signal [invoked from the source thread] > - 52 hits in libc.so`__lwp_park [cond_wait invoked from my code] > - 48 and 41 hits in AES encryption [useful work] > - rest (119 hits) in CV wait/signal and mutex wait/wakeup > > gethrtime() reports total elapsed time of 21.4 seconds. However, > the total > amount of real time spent in top 10 hits is only 872 / 97 = ~ 9 > seconds. > (The rest is probably other application-processing.)
Hrm; try dropping the ', ustack()' from the script; that way, you're only considering kernel stacks, which should raise the amount of signal. > Why is idle() called so often when there's always at least one > thread runnable? > [For if it weren't, the program would never finish.] In both cases, > the idle > time equals ca. 20% of the total real running time (iostat reports > 9% idle -- > I guess that's divided across CPUs). I'm not sure; let's get the kernel stack traces above, and see where it leads. > I tried also to trace the program with plockstat on 8192 threads > (both as > ordinary user and root), and got only the following message: > > plockstat: processing aborted: Abort due to drop You probably need to increase the buffer sizes for this; add - xaggsize=4m to start; if that doesn't help, add -xbufsize=1m, and if that's still not working, make -xaggsize=16m. Cheers, - jonathan > Thanks for your help. > > Best regards, > Zeljko. -------------------------------------------------------------------------- Jonathan Adams, Sun Microsystems, ZFS Team http://blogs.sun.com/jwadams _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
