[ZODB-Dev] Writeperformance for different storages
Barry A. Warsaw
barry@zope.com
Wed, 18 Dec 2002 09:20:47 -0500
>>>>> "TD" == Toby Dickenson <tdickenson@geminidataloggers.com> writes:
TD> I know the BDB documentation says it is a good thing, but did
TD> you measure whether the independant controllers actually made
TD> a difference? to reads or writes? I would not expect that ZODB
TD> (particularly under ZEO) has enough concurrency for this to
TD> make a difference.
I had a dual scsi system for a short while and did a bunch of tests
including separating the log and data files. This was primarily with
the loadmail tests which are more write intensive. It made a huge
difference. I'd need to do some digging to find the exact numbers but
IIRC there was a factor of 5-7 improvement (judicious tuning of other
available parameters may be part of that but not all).
Note that this wasn't with the current cvs of BDB. I don't have
access to that dual scsi machine any more and my feeble hardware
attempts at trying to add a disk to my second IDE channel was less
than successful. If anybody else has the h/w available, I could help
you reproduce the tests to see if we're still seeing that much
improvement.
TD> (If it makes a difference for BDB, I might consider
TD> resurrecting a plan to stripe a DirectoryStorage across
TD> directories on multiple controllers)
Neat.
TD> Its hard to imagine a storage writing any fewer bytes to disk,
TD> or writing them in a way that is any faster to fsync. I think
TD> FileStorage has unbeatable write performance.
Yep.
-Barry