From: Viren Gupta (virender_gupta@TIMEINC.COM)
Date: Mon Nov 04 2002 - 14:24:47 EST
I recently did some performance tests (very unscientifically - because I
was primarily interested in wall clock times of specific apps). The
apps and data used were identical on both systems. Here are some
details which put things into perspective.
System A - S80, 18 processors, PowerPC_RS64-III (450mhz?), AIX 4.3.3 ML 10
System B - p690 12 processors, PowerPC_POWER4 (1.3 ghz), AIX 5.1 ML 2
Both systems are connected to two Sharks with four FC adapters via 2109
switches.
Results
1. CPU bound apps, using Syncsort, gzip and Abinitio ran 30-50 % faster
on system B.
2. I/O bound apps, using home grown C++ code ran about 10% slower on
system B.
Notes.
No surprises with #1, we were quite pleased with those gains.
#2 was a big disappointment. vmtune parameters were set similar.
AIX 5.1 has some additional parms for tuning JFS2, which were tuned too.
Since the filesystems used in system B were JFS2, I decided to
recreate the filesystems in JFS and re-run the tests. Unexpectedly,
we saw improvements of about 25% over system A. This was borne out not
only by wall clock times but output collected from vmtat etc during the
tests (I/O wait %ages as well as number of processes in I/O wait queues).
Has anyone experienced this? I really would like to use JFS2 for its
enhanced features, but the performance was a deal breaker. Oh, one
more issue, about which I have doubts is that systems B is running in 32
bit mode. I have a suspicion that JFS2 with 32 bit was the culprit for
low performance. Will have to wait to test with JFS2 in 64 bit, till
we resolve various incompatibility issues with software/ apps and 64 bit
mode kernels.
Viren Gupta
Time Inc.
This archive was generated by hypermail 2.1.7 : Wed Apr 09 2008 - 22:16:18 EDT