>
> Any idea what causes this? This is a P133, PCI video, 32MB RAM.
>
This just my uneducated guess. Because of the code in goodness() (sched.c)
the current process is favored when calculating the goodness values. I have
played with this quite a bit. I have tried making all processes equal, and
varing degrees of penalizing the current process. On my servers here (fairly
heavy use) they always perform better when I penalize the current process.
My theory as to why it works. A CPU intensive process running on a lightly
loaded machine will eat up its time slices quickly (I beleive each process can
have three time slices before the kernel will reschedule), thus it is penalized
when process are rescheduled. The same process on a heavly loaded machine will
not have a chance to use up its time slices because the process queue is so
long. So when it comes time to reschedule it will most likely "win" (ie become
the next scheduled process). Then then process will continue to "win" because
it is favored by by the goodness() routine, and because it doesn't use up all
of its time slices. This contiuatual winning of one process causes fairly
long "lags".
Try this on your machine and let me know if it helps your situation.
Find this area of code in linux/kernel/sched.c
/* .. and a slight advantage to the current process */
if (p == prev)
weight += 1;
And change:
weight += 1;
to:
weight -= 1;
Recompile, run a few test and please tell me what the effect is.
Remember these are all my uneducated guesses. (some kernel guru is probably
gonna smite me for my attempt)
James
-- James R. Leu Network Administrator CORE Digital Communication Services jleu@coredcs.com