svn commit: r205132 - head/sys/cddl/contrib/opensolaris/uts/common/fs/zfs

[K. Macy]

On Sat, Mar 20, 2010 at 3:10 AM, Pawel Jakub Dawidek <pjd at freebsd.org> wrote:
> On Sat, Mar 13, 2010 at 09:41:53PM +0000, Kip Macy wrote:
>> Author: kmacy
>> Date: Sat Mar 13 21:41:52 2010
>> New Revision: 205132
>> URL: http://svn.freebsd.org/changeset/base/205132
>>
>> Log:
>>   Don't bottleneck on acquiring the stream locks - this avoids a massive
>>   drop off in throughput with large numbers of simultaneous reads
>
> This is another[1] change we agreed (I hoped) to properly evaluate first
> through zfs-code at opensolaris.org.
>

Hi pjd -
I ran it by zfs-code@ a year ago. I never got any meaningful feedback.
I'm not being flippant when I say that I would like you to interact
with them.

One thing you need to also bear in mind is that Opensolaris !=
FreeBSD. I don't mean that in the phk "this is the wrong mailing list"
sense, but rather the two have different behavioral properties. The
deadlock that I had to track down by hand because you had disabled
witness on ZFS locks occurred within 10 minutes on a modest web
serving load and within 2 minutes under fsstress. It had been in ZFS
quite some time when they finally patched it and yet it was a very
serious problem on FreeBSD. Web serving throughput would drop from
2.5Gbps to a few hundred megabits per second without this change. I
can only guess that sx locks do not behave the same as Solaris'
mutexes to such a large extent that we get very different behaviors
and you'll have to keep an open mind to local changes that cope with
that fact.


Thanks,
Kip