2009-10-09 16:45:21

by Tobias Oetiker

[permalink] [raw]
Subject: Linux IO service time - a perfect source of randomness ?

I have just completed extensive benchmarking of ext3 on 2.6.31.2
both on HW RAID6 as well as single Hard Drive. I have written a new
file system benchmark that creates load by reading and writing large
file trees in paralle to simulate a busy fileserver.

The results are as follows:

* as soon as the system comes under strain the service time
for single io operations becomes almost random with huge
outliers of several seconds.

* on HW RAID the best performance is to be gained with
data=ordered and cfq scheduler

* on single HDD setups, only (!) data=writeback with cfq provides
good performance.

* the deadline scheduler did not perform well in any of the
read/write scenarios.

* btrfs has fantastic read performance but does not (yet?) do well
when readers and writers are competing.

my report is on

http://insights.oetiker.ch/linux/fsopbench/

cheers
tobi

--
Tobi Oetiker, OETIKER+PARTNER AG, Aarweg 15 CH-4600 Olten, Switzerland
http://it.oetiker.ch [email protected] ++41 62 775 9902 / sb: -9900