>what kind of slowdown do you get if you set the number of CPUs to a much
>larger value than the number of CPUs available? (this should simulate the
>'no hand-scheduling' case correctly, right?) How many reschedules per
>second do you get in this case (look at vmstat 1 output) - and how many do
>you get if you have hand-scheduling turned on. What i'd like to know is
>the exact cause of the slowdown - scheduler overhead due to large
>nr_running, cache trashing, both, or something else?
>
A very bad slowdown -- it's not just the scheduling and cache issues.
Each thread does some dynamic memory allocation -- if I let a very
large number of threads run concurrently I run out of RAM and start
swapping. I tried running with the number of CPU's set to a large
value to test scheduler/cache issues but I can only get to 100 or
so on the Linux boxes we have before swapping starts. I do see some
effects but nothing dramatic.
Emil
-
Linux SMP list: FIRST see FAQ at http://www.irisa.fr/prive/mentre/smp-faq/
To Unsubscribe: send "unsubscribe linux-smp" to [EMAIL PROTECTED]
