On Thu, 2008-11-06 at 20:41 +0100, Pau Garcia i Quiles wrote: > The disk IO performance decrease from Gutsy to Hardy is anything but > anecdotal. > > This ( > http://groups.google.com/group/zumastor/browse_thread/thread/7e413960ddc22811# > > ) bug report in the Zumastor project has some (quite scarce) info, > although if you read the IRC logs, you'll realize how much pain it > caused (it made Zumastor unusable due to slowness). It seems that the > scheduler included by default since kernel 2.6.23 cuts performance in > half for certain IO operations. > > Same thing for the CPU scheduler. For instance, here VMware Server is > barely usable when high disk IO is requied because for some reason, > after a few seconds of good performance (which implies high CPU usage) > the kernel starts throttling CPU and VMware starves for a few seconds. > What surprises me about this CPU throttling is it achieves nothing: > VMware is left with 10% of CPU and the other 90% is idling! > Remember that you can change the scheduler on the fly.
The default scheduler is optimised for general desktop usage, where you have a large number of simultaneously running applications, applets, etc. and each one needs to be responsive. It performs badly at single operations that wish to consume all of the CPU or IO resource available. That includes disk copies, VMware, and funnily enough - benchmarks ;P The scheduler would fair extremely well if you compared, say, 20 simultaneously running benchmark suites between earlier releases and this one. All 20 would show fair results. I've always thought it would be interesting to be able to influence the scheduler on a per process basis - and do that from the Window Manager. ie. deliberately give the user's foreground process the "majority" of the time, and fair schedule the rest. > I'm forward-porting kernels 2.6.20 (Feisty's) and 2.6.22 (Gutsy's) > to > Hardy to verify it is the scheduler what is causing trouble. It will > be available in my PPA ( http://launchpad.net/~pgquiles/+archive ) > tomorrow, in case anyone is interested. > Why do you need to forward-port? The same kernel binary will just work. Also you can just fiddle on a per-disk basis, e.g.: echo -n deadline > /sys/block/sda/queue/scheduler Scott -- Scott James Remnant [EMAIL PROTECTED]
signature.asc
Description: This is a digitally signed message part
-- Ubuntu-devel-discuss mailing list Ubuntu-devel-discuss@lists.ubuntu.com Modify settings or unsubscribe at: https://lists.ubuntu.com/mailman/listinfo/ubuntu-devel-discuss
