[nSLUG] performance and schedulers
peter at cordes.ca
Thu Jun 23 18:36:32 ADT 2005
On Wed, Jun 22, 2005 at 08:19:21AM -0300, Jim Haliburton wrote:
> Good Day all:
> I have read the various items on schedulers and performance and have several
> 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?
> 2) Would a selection of higher performance components in the system give you a
> quicker and cheaper performance gain? In other words is it more productive to
> just go buy a bigger box? I am more hardware oriented and find that well
> matched components often give much better performance gain, than hours spent
> tweaking the OS. Thoughts??
Some things are limited by OS algorithms, other things are limited by
hardware. Running one CPU bound program is limited by hardware; Even the
most stupid OS process schedulers can't screw that up.
As for tunables, with 2.6 I found that on a desktop machine running
bittorrent and experienceing some serious jerky behaviour and swapping out
mozilla every time I looked the other way, turning down the swappiness
helped a lot. Bittorrent can be running all the time and you don't notice
the effect on the rest of the system. /proc/sys/vm/swappiness is the main
VM tunable in Linux 2.6. It was at 60, and I turned it down to 20 or 30 IIRC.
One a machine at work with 3Ware RAID, 3w's docs recommend setting
readahead way higher than it defaults to. They seem to be describing this
as the way to do well on a sequential I/O benchmark, though. They were
to run blockdev --setra 16384, but I only used 4096. (I have a 4-disk
RAID5; maybe a bigger array could reasonably gain from something ridiculous
That's about all I know or remember about kernel tuning at the moment. I
haven't experimented with different I/O or process schedulers. Hopefully
sometime I'll find a patch for batch-scheduling that works with a kernel I
want to run. That way my Folding at Home processes won't get any CPU when
anything else wants it.
Oh, and since this is my first post in a long time, I should mention that I
haven't been keeping up with my nslug email. I haven't even been lurking...
#define X(x,y) x##y
Peter Cordes ; e-mail: X(peter at cor , des.ca)
"The gods confound the man who first found out how to distinguish the hours!
Confound him, too, who in this place set up a sundial, to cut and hack
my day so wretchedly into small pieces!" -- Plautus, 200 BC
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 351 bytes
Desc: Digital signature
More information about the nSLUG