> You do not need to follow exactly this procedure to observe the
> performance loss. You just simply launch 1 instance of prog1 when you
> just have rebooted your system, observe the timing, and then launch a
> large number of them. Then kill them all and launch one more instance
> of prog1: now prog1 should take more time to complete and make
> overruns.
Here, on a K6-2/300 with 2.1.120, it does not work like you describe :
First run :
150 iterations in 34.956501 seconds with 349 overruns
150 iterations in 35.211155 seconds with 352 overruns
150 iterations in 35.028309 seconds with 351 overruns
150 iterations in 34.935802 seconds with 349 overruns
Second run (50 prog1 launched in // and killed) :
150 iterations in 30.217871 seconds with 301 overruns
150 iterations in 29.913172 seconds with 299 overruns
150 iterations in 29.746883 seconds with 297 overruns
150 iterations in 29.875475 seconds with 299 overruns
Third run (again 50 prog1 laucnhed in // and killed) :
150 iterations in 32.790020 seconds with 327 overruns
150 iterations in 32.836475 seconds with 329 overruns
150 iterations in 32.832195 seconds with 328 overruns
150 iterations in 32.770821 seconds with 328 overruns
It looks like a cache colouring problem.
Hint : disable the L2 cache and see if the problem persists.
BTW, Linux has very bad interactive performance when the load is high. IIRC,
it was much better in the past. I also tested prog1 on a Solaris 2.6 box, and
it was MUCH more responsive with a load above 100 than the Linux box with only
one prog1 running.
-- Benoit
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/faq.html