[nSLUG] performance and schedulers
donald.teed at gmail.com
Thu Jun 23 07:44:31 ADT 2005
On 6/22/05, Bill Davidson <bdavidso at supercity.ns.ca> wrote:
> I'm not sure you are right there, or at least your kernel version is a
> bit off. I think at least some of the change came in 2.6.10, when the
> schedulers were made modular -- that is, you can compile them as
> loadable modules.
OK, I just knew it appeared after 2.6.8 - so "thereabouts".
Finding out which one you are currently using is done via the sysfs
> bdavidso at plato2:~$ cat /sys/block/hda/queue/scheduler
> noop [anticipatory] deadline cfq
> Note that "anticipatory" has "" around it, indicating that it is
> active for that device. You can change it thus (from the 2.6.10
> Changing is done with
> router:/sys/block/hda/queue # echo deadline > scheduler
> router:/sys/block/hda/queue # cat scheduler
> cfq [deadline] anticipatory
> deadline is now the new active io scheduler for hda.
> This really ought to be documented better somewhere -- searching through
> a 1.5M text file is no fun.
That's assuming you know it is documented in that changelog.
I was looking for info under the kernel's Documentation dir.
Kinda cool though, eh? You can have different schedulers for different
> devices. I assume this would only apply to low-level IDE devices, and
> not to md devices. What do you see under /sys/block/<device>/queue on
> your RAID system?
It is cool and nearly radical! Can you imagine an airplane that can
change to wooden propellers from metal propellers in mid flight while
the engine still spins?
I was going to make another 2.6.12 kernel last night to run
the benchmarks against md raid with different schedulers
as you suggested, but I'm running into something that
prevents full boot up.
I get an error that the superblock is bad on every raid 5
and that fsck should run. But /dev is not getting built
properly by udev (which was working before)
so the problem is somewhere between that and some
kernel option I've messed up. I'm pretty sure the raid
partitions are fine, because 2.4 boots with it OK.
I know that udev needs hotplug, and I know I didn't
need to make rules for it before, so it is something else.
Maybe some day you'll be able to cat vmlinux > /sys/kernelimage
to switch kernels on the fly!
It drives me nuts when something is required, the
usual way to handle it becomes depreciated, and
the new way is still on the edge of experimental (they
admit it won't work as well as devfs for some devices).
I like the way they say that the devfs developer
"disappeared". Visions of Anti-Trust.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the nSLUG