Re: Linux 2.6.15-rc5: multiuser scheduling trouble

[Helge Hafting]

2.6.15-rc5 (and earlier 2.6 kernels) seems to have a slight scheduling 
problem, with some starvation at a load of only 2.

All it takes is two users with separate X desktops.
User one plays flash games using firefox. The ill-designed flash stuff
typically means that his Xorg and firefox divides the cpu 100% between them.

So I expect that I, as user two, should notice some slowness. I should get only
1/2 - 1/3 cpu.  But I get periods of starvation.  Logging in takes a long time,
bringing iup icewm takes 15s instead of 2, each xterm takes a long time to
start.  They are usually instantaneous.  Tha machine is unsuitable
for work in this mode.

Knowing the root password I renices his Xorg and firefox by 10, and then
everything is fine.  His games are still ok, and my xterms are snappy again.

I have tried no preempt, voluntary preempt, and preemptible kernel.  It doesn't
make a difference.  This is an amd64 kernel on an opteron 244 (1800MHz).  Everyhting
is 64-bit except firefox+flash which is 32-bit.

Perhaps the way flash games work, with lots of communication with the xserver,
makes them get "io boost" even though they are cpu hogs. I still think
my xterm (or whatever I am starting up) should get its fair third of the cpu
though, (with firefox and xorg hogging one third each too.)  Even a "600MHz opteron"
ought to do better than this.

The machine isn't trashing, it is hardly touching swap.  (512M memory, and swpd=16 
according to vmstat) The paging-in of a starting executable shouldn't be affected much
by the cpu load?

Helge Hafting
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/