On Tue, 2005-01-18 at 23:42, Yusuf Goolamabbas wrote:
> > Why did you reduce TCQ Depth ?
> Nick Piggin has mentioned it on lkml that AS with high TCQ doesn't work
> well. For Database servers, it's recommended to use deadline. See
> Documentation/block/as-iosched.txt in the linux-kernel source tree. I
> got similar numbers with deadline and default tag depth of 32
Yes. Deadline seems to be best for database workloads. However as I
mentioned in your case it is not likely to be disk bound at all.
> I'll try to setup a RAID 10 disk mirror and then re-run the same tests
> with deadline/cfq/as.
> > Yes. 2.6 kernel seems to do much better. It is however barely depending
> > on IO subsystem as default table size is just 10000 rows so it will well
> > fit in memory.
> I changed the table size to 100000 rows. If/When I get a faster i/o
> system on the boxes. I'll try with even larger table sizes.
Well it is still just 250MB or so. The rows are some 200 bytes if I'm
not mistaken. You need to be significantly larger than memory on the
box to be really disk bound. Further more distribution is very skewed
by default so most accesses will come from even smaller space.
Also make sure you do not repeat requests between runs if they are short
enough (use init random generator from timer option)
Peter Zaitsev, Senior Support Engineer
MySQL AB, www.mysql.com