tech-kern archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: mpt Serious performance issues

On Wed, 22 Dec 2010, Stephan wrote:

Does anybody have an idea what causes the poor perfomance on hardware

  Write performance can be significantly reduced due to lack of write
caching in some raid adapters.  Typically, raid adapters will require
some kind of battery backup in order to enable write caching.  The only
adapters I'm very familiar with in this regards are the Compaq/HP Smart
Array adapters (ciss(4)).  The Smart Array adapters will not allow you
to enable write cache if there is no battery, but I have been able to
enable write cache on a Dell M600 Blade with mpt(4) by using a linux
lsiutil program.  The adapter on the M600 does not have battery backup,
so there's more potential for data loss on loss of power.

I saw the same with the ciss driver which is as slow on 5.1 but very
fast on 5.99.39 (up to 180MB/s write speed).

The changelog tells about "tagged queueing"...

  I found out through experimentation that the Smart Array adapters
will perform much better with no write cache if multiple commands are
queued simultaneously to the adapter.  Due to the way NetBSD queues
command via the scsipi mid-layer, and the ciss(4) driver not notifying
the mid-layer that it can accept multiple command, ciss(4) performed very
poorly until I was able to fix it.

  The mpt(4) driver has code that attempts to enable tagged queueing, but
as best I can tell, there isn't any way it could ever be enabled.  I'm
also not at all certain that the mpt adapter performance would improve
by queueing multiple commands, and I don't have a system at this time to
do any testing on.  I was working with someone with a Dell system , but
she has been rather involved in other things recently.

  If I can ever get access to the M600 blade again, I can try to figure
out how to make the driver enable tagged queueing and see if that makes
any difference, but that blade is currently being used for other testing
with a much highor priority than my playing around.

Michael L. Hitch              
Computer Consultant
Information Technology Center
Montana State University        Bozeman, MT     USA

Home | Main Index | Thread Index | Old Index