Report of performance on Alpha vs. Intel - Mailing list pgsql-general

   This week, I had the opportunity to compare the performance of PostgreSQL
on an Alpha and an Intel server, and the results kind of surprised me.  I'd
love to hear if this has been the case for others as well...

-------------
Intel Machine

SuperMicro 8050 quad Xeon server
512 MB RAM
4 x PII Xeon 400 MHz (secondary cache disabled)
RAID array w/ 5 9-gig drives

Approximate cost:  $6000
--------------
Alpha Machine
AlphaServer DS20E
2 x CPU (500 MHz or 667 MHz)
2 GB RAM
9-gig SCSI drive

Approximate cost:  $20,000 - $25,000
-----------------------

General System notes

    I'm not sure which chips the Alpha uses, the 500 MHz or the 667 MHz.
Also, because the SuperMicro board is meant for the newer Xeons, the
secondary cache had to be completely disabled on the PII 400 Xeons, so that
machine was definitely not running up to potential.

-------------------------
Test method

   This wasn't exactly the ANSI tests, but it accurately reflected what we
need out of a machine.  A while back we logged 87,000 individual queries on
our production machine, and I selected one thousand distinct queries from
that.

   On each machine I spawned 20 parallel processes, each performing the
1,000 queries, and timed how long it took for all processes to finish.

   To try and keep the disk subsystem from being a factor, this used only
selects, no updates or deletes.  Also, the database is small enough that the
entire thing was easily in the disk cache at all times.
--------------------------
Test results

  The Alpha finished in just over 60 minutes, the Xeon finished in just over
90.

-----------------------------
Test interpretation

  Once I started looking at the numbers, I was suprised.  On a
processor-for-processor basis, the Alpha was three times as fast as the
Intels.  However, the Intels that it was pitted against were only 400 MHz
chips, only PII (not the PIII), *and* had the external cache completely
disabled.

   So, the Alpha provided three times the performance for four times the
cost - but if the megabyte of cache had been enabled on the Xeons, I think
that the results would have been significantly different.  Also, if the
chips had been even relatively recent chips (say, some 700 or 800 MHz Xeons)
with the cache enabled, it's possible that it could have come close to the
performance of the Alpha, at a much lower cost.

  Overall, I was expecting the Alpha to give the Intel a better trouncing,
especially considering the difference in cost, but I guess it's hard to beat
Intel for transactions/dollar.  If sheer server capacity is the only
relevant factor, forget Intel (You won't find Intels with 64 processors, and
I don't think you'll see them even with the Itaniums).  If your needs are
more down-to-Earth, they're the best you can get for the money.

steve