Volker Armin Hemmann wrote: >Hi, > >On Wednesday 30 March 2005 19:07, Pupeno wrote: > > >>Je Mardo Marto 29 2005 13:59, Volker Armin Hemmann skribis: >> >> >>>On Monday 28 March 2005 09:51, Pupeno wrote: >>> >>> >>>>PORTAGE_NICENESS is set to 19 on /etc/make.conf. >>>>Any ideas ? >>>> >>>> >>>yes, stop using 'niceness'. >>> >>>With 2.6 nice does not work anymore like it worked with prior releases. >>>Processes that are niced to negative values might get a lot less CPu than >>>they should, while positive values get a lot more CPU cycles than they >>>should. >>> >>> >>Is this documented somewhere ? >>Thanks. >> >> > >yes, but emm, I do not have any links ;) > >The 'problem' is, that 2.6 favours 'nice' processes and penalises CPU-hungry >or 'not nice' processes. This is why you should not renice X with 2.6. > >
I think this is almost completely false. Or at least misunderstood. According to Documentation/sched-design.txt, the user specified priority (i.e, nice value) still applies. The kernel will penalize a CPU-hungry process by lowering it's priority (increasing it's niceness), but it only goes so far. I remember reading somewhere that it maxed out at +5, but maybe this has been adjusted too. We can check the kernel sources to be sure. So, you are correct in that a mostly-sleeping process at nice 1 can get scheduled before a cpu-hog at nice -1. But the nice value is still taken into account, particularly in relation to other cpu-hogging or cpu-friendly processes. So the bottom line is that emerging with niceness 19 is still useful...it tells the scheduler to "run this thing last". It might not be as critical as it was with 2.4, but it still has the exact same effect. -Richard -- [email protected] mailing list
