At 12:42 PM 3/6/2003 -0800, Andrew Morton wrote:
>Ingo's combo patch is better still - that is sched-2.5.64-a5 and your patch
>combined (yes?).  The slight mouse jerkiness is gone and even when doing
>really silly things I cannot make it misbehave at all.  I'd handwavingly
>describe both your patch and sched-2.5.64-a5 as 80% solutions, and the combo
>95%.

Weeeeell, FWIW my box (p3-500/128mb/rage128) disagrees.

I can still barely control the box when a make -j5 bzImage is running under 
X/KDE in one terminal and a vmstat (SCHED_RR) in another. I'm not swapping, 
though a bit of idle junk does page out.  IOW, I think I'm seeing serious 
cpu starvation.

For reference, I used to be able to easily control the same box under X 
while swapping moderately at make -j15 (during page aging vm days), and 
with a heavily twiddled VM at even -j30 (_way_ overloaded with X oinkers 
running).  At -j5, gui/icky was quite usable.

With the make -j30 bzImage, heavy load which does _hit_ VM saturation but 
not constantly, I see what I can only describe as starvation.  vmstat 
output for this load is definitely abby-normal, and there's no way 
/proc/vmstat should look like the attached.  It takes forever to build 
enough vm pressure that swap is touched (bad bad bad), and the volume 
reached is way low.  My hungry little cc1's are not sharing resources the 
way they should be (methinks).

With earlier versions of Ingo's patch, I was only able to get memory 
pressure up to where it should be by drastically reducing timeslice (at the 
expense of turning throughput to crud).  Take this with a giant economy 
sized grain of salt, but what I _think_ (no, I haven't instrumented it) is 
happening is that chunks of the total job are being rescheduled in front of 
the rest of the job such that they alloc/use/exit before the other guys can 
alloc.  (geez that's a nebulous WAG;)

         -Mike  (please torch me if I'm being stupid;)