[nSLUG] performance and schedulers

Stephen Gregory nslug at kernelpanic.ca
Wed Jun 22 13:36:19 ADT 2005


On Wed, Jun 22, 2005 at 08:19:21AM -0300, Jim Haliburton wrote:
> 
> 1)  With a careful test of each of the various schedulers and their tunable 
> paramters, how much performance gain could one expect?   Are we talking 10. 15 
> or 30% change?

In most cases it will only be a few percent change in performance if
any. Until you hit a full load condition with many processes and
threads. At which point a kernel with the correct scheduler will work
while another kernel wont. The different schedulers have been added to
solve a specific problem or provide a gain in a specific case. The
performance gain is always a tradeoff of one performance metric for
another.

The prempt schedulers were added to give better response time to
workstation users. It does work and is well tested. It is a little
tricky to benchmark as it actually hurts performance to give an
allusion of speed to the console user.

There isn't much point in an routine that switches schedulers. Such a
routine would have to optomize for some performance metric. It is
easier and faster to just write a scheduler that optomizes for that
same performance metric.

-- 
sg


!DSPAM:42b9938e62571857312814!




More information about the nSLUG mailing list