On Sun, 8 Nov 1998, Alan Cox wrote:
[about PROC_CHANGE_PENALTY processor-affinity]
> > Setting this to eg. 100 should make the process less prone to
> > switching CPU.
> Its actually very tricky to tune since it really depends on the machine
> and process properties. The current algorithm is probably too simplistic
> but it worked on my dual P90 ;)
later kernels already have a mechanizm to do some more scheduling on SMP,
do an #if 1 at line 113. It's not completely safe yet, but it will be an
interesting 2.3 item to work out. (or custom patch)
can anyone write some code that clearly demonstrates that we lose
performance on SMP? The line 113 code helps mostly latency, which i can
demonstrate to make a difference, but it looks quite hard to demonstrate
cache-affinity problems. (have tried and failed, if someone is able to
write code that demonstrates how we lose performance on a dual-PII system,
i'd be very happy to fix the problem)
-- mingo
