From: Tim Chipman (chipman@ecopiabio.com)
Date: Fri Jan 10 2003 - 10:06:49 EST
Hi All,
A bit of a tuning question which I'm hoping to get some insight into.
Searches of google and the list archives reveals nothing beyond what I
knew before starting this "performance enhancing tweak" for our oracle
server, which appears to be having unexpected consequences.
We've got an e450 (quad 400/2gig ram/Internal SWraid for OS slices, T3
disk array for Ora Datafile slice) running Solaris8 and Oracle 8i as our
Database server.
Up until recently, all filesystems were vanilla solaris-UFS with logging
enabled, NO other options set.
For a while, I had been planning to tweak the mount options on the
"oracle slice" by adding these options into the vfstab for this slice:
forcedirectio, noatime
This change was made ~2 weeks ago (stopping oracle ; unmounting OraData
slice ; remounting with new settings ; restarting oracle), and since
that time, there has been a distrubing trend: Observation with "top"
indicates that at times of "significant database activity" (ie, daytime
hours), the "baseline" levels of IOWait are MUCH higher than previously.
ie, in the past, iowait might spike to 50-60% for a few seconds, then
drop back to ~0, more-or-less as large queries came in to oracle.
What I'm seeing now is sustained periods of IOWait above 50%, often
upwards to 75%. This is starving the CPU resources available to Oracle ;
the queries themselves are going nowhere since oracle processes
themselves can only get 1-2% CPU resources instead of 10-20% for these
"query bursts", and the situation is moderately grim.
I've really confirmed this by looking back at our MRTG graphs logging
system load behaviour ; there is a clear change in the baseline "system"
process loading from <generally less than user process loading> to
<generally more than user process loading>. The behaviour change is
exactly aligned with the change to mount options for this OraDataSlice.
I'm curious if anyone has any comments on this or has observed similar
behaviour -- or the like. "Dogma" suggests that forcedirectio should
have made oracle much happier (since oracle caches plenty for its own
purposes, hence prevention of the OS from meddling is a "good thing" in
theory?) but this seems to be causing other problems, in a big way.
Thanks everyone for your help, and as always, a summary will follow ...
--Tim Chipman
_______________________________________________
sunmanagers mailing list
sunmanagers@sunmanagers.org
http://www.sunmanagers.org/mailman/listinfo/sunmanagers
This archive was generated by hypermail 2.1.7 : Wed Apr 09 2008 - 23:25:34 EDT