New scheduler algorithm committed
dillon at apollo.backplane.com
Sun Jun 26 20:10:57 PDT 2005
:My very preliminary testing shows an improvement, but not spectacular.
:I'm doing a buildworld now, just as a CPU-intensive background task.
:I've tried playing the gnome 'sol' game, which seems definitely more
:responsive. I'm typing this response in pine (running in a gnome
:terminal window) and I find it so annoyingly slow and erratic as to be
:So, all in all, the scheduler changes seem to be a tossup: some pluses
:and some minuses. Still room for improvement, though.
Ok, I've set a machine up with gnome and pine and I see a few issues.
The main one is that the estcpu decay time calculations are occuring
on one-second intervals without any regard as to whether the process
has even existed for one second or not. buildworlds with their very short
lived processes are being improperly accounted for. This is something
we inherited. I'm going to fix that, do some more testing, and hopefully
have something by tonight.
gnome is a good test. The gnome-terminal program is a medium-sized cpu
hog (it uses four times the cpu that the pine program running under it
uses). I think the system is accounting for the program properly, but
is skewing the buildworld due to the estcpu decay algorithm.
<dillon at xxxxxxxxxxxxx>
More information about the Kernel