Help with improving mysql performance on 6.2PRE

[Sven Petai]

On Thursday 05 October 2006 03:52, Jerry Bell wrote:
> I expected the 2950 to be a bit closer to the 1.8Ghz opteron discussed
> here:
> http://article.gmane.org/gmane.os.freebsd.performance/1137/match=mysql

Well he used several patches [1] from the current + ULE scheduler, which seems 
to give you better results in this particular benchmark, but sometimes seems 
to have some weird sideefects for realworld usage. Also keep in mind that 
default select-key smack is just very hard on some specific subsystems. The 
load it produces is rather unrealistic so better mysql supersmack scores 
might, but don't necessarily, translate into better real performance.

You might get some idea on how different patches and kernelconfigs affect 
supersmack results on SMP from my latest benchmark run on 8 core machine:
http://bsd.ee/~hadara/debug/mysql_sep_2006/stats.html

With 8 cores my performance results are of course seriously pessimized by 
locking contention, I'm not sure how much of a problem it's with 4 cores, but 
it's very likely that rwatsons uds locking patch [2] that I used will help 
some in your case too, though I have no idea if it applies to 6.X tree 
without some hacking.

-------
[1] 
http://lists.freebsd.org/pipermail/freebsd-performance/2006-April/001869.html
[2]
http://www.watson.org/~robert/freebsd/netperf/20060822-uds-fine-grain.diff