seti@home and nice

cmay4 · Tux's lil' helper Joined: 18 May 2002 Posts: 109

I thought that if something was niced to 19, then it would essentially idle if other things were using the cpu. This doesn't seem to be the case. Look at the top output:

sputnik1969 · Posted: Tue Apr 01, 2003 12:02 am Post subject:

cc1plus CAN'T use all cycles, because it has sometime to wait for fileoperations and similar. In this moment the scheduler switches to setiathome and it's getting "his" part of the CPU for a while.
_________________
'Cynic' is a word invented by optimists to criticize realists. - Nigel. In the ocean of Night by Gregory Benford
"Zyniker" ist ein Wort das Optimisten erfunden haben um Realisten zu kritisieren. - Nigel. Im Meer der Nacht von Gregory Benford

cmay4 · Tux's lil' helper Joined: 18 May 2002 Posts: 109

sputnik1969 · Posted: Tue Apr 01, 2003 3:29 am Post subject:

cmay4 · Tux's lil' helper Joined: 18 May 2002 Posts: 109

I'm sure it was my misunderstanding. I was just under the impression that something with that nice value would not (or at least barely) affect other running software. I made a slight modification:

sputnik1969 · Posted: Tue Apr 01, 2003 4:46 am Post subject:

cmay4 · Tux's lil' helper Joined: 18 May 2002 Posts: 109

Kernel compile numbers:

With seti@home: 13min 30sec
Without seri@home: 11min 32sec

Or 14.5% slower with seti@home. 5% is one thing, but that is too much.

Doesn't that seem to be too large a difference?

nempo · Posted: Tue Apr 01, 2003 2:18 pm Post subject:

Everytime I've had seti@home running I've been unable to run any kind of opengl software, why? well every 2s or so the rendering stops for ½s and then continues, this constantly happens. Seti@home was running at 19 while the opengl ran at the default 0.
_________________

sputnik1969 · Posted: Tue Apr 01, 2003 2:46 pm Post subject:

Everything here are based upon the same problem:

Linux doesn't know "true" idle-processes and the priority-system isn't "fine" enough...
Multiprocessing under Linux works with "timeslices" / slots.
If there are 2 processes running with different priorities (e.g. process Seti Prio 19 and Q3A Prio 0) both processes share the slots, seti gets at least minimal 1 slot of 20 (difference between priority of Seti and Q3A) and Q3A gets the rest... But this single timeslice has the effect, that 1/20 (5%) of the time quake has to wait for seti, even if seti should be idle

If priority-system would range from -128 to 127 (like Amiga-OS) and Seti would run with 127 and Q3A with 0, it would get at least 1/128 (<1%)
But then timeslices have to be shorter, or latency will be a really big problem

Also this needs more cputime for the scheduler

_________________
'Cynic' is a word invented by optimists to criticize realists. - Nigel. In the ocean of Night by Gregory Benford
"Zyniker" ist ein Wort das Optimisten erfunden haben um Realisten zu kritisieren. - Nigel. Im Meer der Nacht von Gregory Benford

cmay4 · Tux's lil' helper Joined: 18 May 2002 Posts: 109

Thanks for the input. I appreciate your help.

I have one question though. I've read about the new scheduler in the 2.5 kernel. Would it make any difference in this situation?

zephyr1256 · Posted: Tue Apr 01, 2003 8:01 pm Post subject:

I have to turn Seti off for opengl apps and for one other case(that I've found so far), whenever I use openoffice.org, seti has to be turned off. Often OO won't even display if I switch to the screen that has it if seti is running.

Nice levels don't help with this, I always use 19, but Seti, unlike most other programs is highly CPU bound. I guess some programs often do system calls but they depend on getting the CPU again sooner than they do with Seti running. Programs like that need to have Seti shut down(or relegated to another processor if you are fortunate to have a MP machine).

sputnik1969 · Posted: Fri Apr 04, 2003 2:41 am Post subject:

sputnik1969 · Posted: Fri Apr 04, 2003 2:53 am Post subject: