From: Sheldon Lee Wen (sheldon.leewen@cgi.com)
Date: Tue Jun 10 2003 - 10:23:01 EDT
Hi,
We have Oracle setup and installed on a virtual partition on a SAN. Oracle
is doing direct IO, and uses it's own buffer cache. We are on an ES45, 4GB
RAM, dual processors, Tru64 5.1A
Over a 20 hour period, Oracle performed almost 29 million reads and 6
million writes over 20 hours. Most of these (less a million or so that were
multi block) were single 8K Oracle block reads (probably lots of index
reads). The numbers over 20 hours work out to about 480 I/Os / second, 3.8
megabytes / second for the entire 20 hours.
Oracle reports 5ms on average for single block I/Os, while its multi block
I/Os are a bit longer (14ms for a 128 block / 1 meg read), but a lot more
efficient (128 blocks in only 3 times the response time).
The SAN should be able to push 238 meg/sec (2Gb fiber), so we were not even
close to that. However, Oracle is performing a very large number of small
I/O requests, perhaps at a high enough average rate to saturate the I/O
queues, causing excessive waits? Do you guys have any idea of what kind of
I/O rate numbers the O/S and SAN (HS110) can sustain? Are we hitting a limit
somewhere? Is there a monitoring tool that shows queue depths (or
something) that is easily available w/o additional licensing?
Thanks
Sheldon.
This archive was generated by hypermail 2.1.7 : Sat Apr 12 2008 - 10:49:21 EDT