On Thu, 14 Jun 2007, Sunny Bains wrote:
> Jeff,
>
> Jeff Roberson wrote:
>> Firstly, has anyone examined using pthread standard locks in place of
>> your home-rolled spinlocks?
>
> I assume you mean the InnoDB mutex implementation. I've tested the pthread
> locks under Solaris 10 and it does help there. I'm interested in your
> observations and your comments are most welcome.
I have produced a graph of mysql writes on linux and bsd:
http://people.freebsd.org/~jeff/mysqlwrite.png
This is using the sysbench oltp benchmark. I see long periods with no
disk activity and almost no CPU activity where all mysql threads are
blocked on a condition variable. This can last for a second or two. Has
anyone investigated it? Is this an understood problem?
The difference between the blue and green lines on FreeBSD is a patch to
reduce kernel contention related to fcntl(). However, the most
significant sources of contention by far are in mysql.
Thanks,
Jeff
>
> Regards,
> -sunny
>
