At 03:15 PM 11/26/2005, Luke Lonergan wrote:
I suggest specifying a set of basic system / HW benchmarks to baseline the
hardware before each benchmark is run. This has proven to be a major issue
with most performance tests. My pick for I/O is bonnie++.
Your equipment allows you the opportunity to benchmark all 5 machines
running together as a cluster - this is important to measure maturity of
solutions for high performance warehousing. Greenplum can provide you a
license for Bizgres MPP for this purpose.
...and detailed config / tuning specs as well for it or everyone is
probably wasting their time. For instance, it seems fairly clear
that the default 8KB table size and default read ahead size are both
pessimal, at least for non OLTP-like apps. In addition, there's been
a reasonable amount of evidence that xfs should be the file system of
choice for pg.
Things like optimal RAID strip size, how to allocate tables to
various IO HW, and what levels of RAID to use for each RAID set also
have to be defined.
The 16x SATA drives should be great, provided you have a high performance
RAID adapter configured properly. You should be able to get 800MB/s of
sequential scan performance by using a card like the 3Ware 9550SX. I've
also heard that the Areca cards are good (how good?). Configuration of the
I/O must be validated though - I've seen as low as 25MB/s from a
misconfigured system.
The Areca cards, particularly with 1-2GB of buffer cache, are the
current commodity RAID controller performance leader. Better
performance can be gotten out of HW from vendors like Xyratex, but it
will cost much more.
Ron
