Anticipatory vs Deadline vs CFQ

kam1su2 · 04-14-2007, 11:05 PM

Trying to figure out which I/O Schedular is best for every day usage, from playing music to watching videos... all the way to running small databases. I hear mixed opinions... does anyone know the low down exactly?

bird603568 · 04-14-2007, 11:10 PM

i use cfq but that doesnt mean its the best

H_TeXMeX_H · 04-14-2007, 11:59 PM

I'm betting either deadline or CFQ. I use deadline. Try 'em out and see.

kipp · 04-15-2007, 05:38 PM

If my memory servers me right, the kernel help on the 3 different schedulers states the CFQ should be suitable for most dekstops. The Deadline and Anticipatory helps do not mention desktop in their help description, and is why I chose CFQ over them. As far as any real benchmark of it, I can't provide that.

bird603568 · 04-15-2007, 06:21 PM

in real life i dont really think its matters like i dropped the timer frequency from 1000hz to 100hz and i dont notice a differnce

H_TeXMeX_H · 04-15-2007, 06:24 PM

Here's more on the schedulers ... how they work and how to switch between them right now if you want.

http://www.linuxhowtos.org/System/iosched.htm

syg00 · 04-15-2007, 06:51 PM

The big commercial distros have been using CFQ for a while. Redhat had a (white ???) paper on their testing up on their site somewhere.
One of the recent kernels has changed the default to be CFQ. (2.6.18)

syg00 · 04-15-2007, 07:01 PM

Dup - deleted

H_TeXMeX_H · 04-15-2007, 07:08 PM

If anyone finds some nice benchmarks on this, pls post it ... I haven't been able to find any yet. (except for one from oracle ... can't say I trust them tho)

bird603568 · 04-15-2007, 10:16 PM

i dont know this is pretty hard to benchmark. all that it is is the order of stuff that gets executed. i would assume that deadline would get stuff with a higher nice run first

H_TeXMeX_H · 04-15-2007, 11:23 PM

Well, it seems to be a bit more complex ...

Here's a benchmark from Red Hat (2005):
http://www.redhat.com/magazine/008ju...es/schedulers/

Also, it seems to depend on your filesystem (as well as workload):
http://www-941.ibm.com/collaboration...I%2FOSubsystem

Code:

Selectable I/O Schedulers (2.6 feature only)

    * workload & FS dependent
    * CPU bound workload - use noop I/O Scheduler
    * non-CPU bound workload - CFQ and deadline
    * JFS works better on CPU bound workload

I use JFS, so I guess I should be using noop. Maybe I'll try it out.

syg00 · 04-16-2007, 04:08 AM

Mmmm - I hope you re-read that reference.
- are you ppc ??? (IBM proprietry RISC chip)
- are you CPU bound ??? (in the p-series case probably 100% on more than 1 core)
- are your I/O being delayed because of *all* of the above ???.

jfs is very much the after-thought in that recommendation - that wiki is very specific to IBM.

I doubt anybody outside of significant enterprise datacentres could pick the difference in the I/O schedulers unless they have a specific need.
Last time I looked at the code I decided it wasn't even worth testing for. I had thought noop might show some benefit where hardware RAID was in play, but couldn't see any real-world situation where it would be worthwhile testing.

H_TeXMeX_H · 04-16-2007, 10:52 AM

Yeah, well, I guess just try them out ...