hyperthreading, nice level problem
 

Hi,

I just benchmarked mprime again on a P4/HT with smp/smt kernel.
The results were as I expected - The "mprime throughput" is about 10% higher if I run 2 instances at the same time.

But by the way I noticed something odd:
If I run two mprimes at the same time - one with nice level 19 and one with 0 - then both instances are needing the same time per iteration.
So the nice mprime harms the other mprime - setting niceness becomes useless.

For me this seems to be logical on a HT CPU since the scheduler gives both active processes 100% of one CPU (which are only a virtual CPUs!).

My problem is clear now, even if I run only one nice mprime on that machine - it steals about 50% CPU time when another (more important) process becomes active.

Is there any solution for that exceptly disabling HT/smp?
My kernel is 2.6.13 - maybe the HT support improved since then?

For now I will play around with MaxLoad=1.5 in prime.ini but its not really satisfactory because mostly mprime is still not sleeping even another process is active since 1 or 2 minutes.

You can tell mprime* to stop computing once certain processes are running. AFAIK, the required option is in undoc.txt.


* Actually, I only know for sure that prime95 is capable of this, but I guess it's similar with mprime.