Steve,
Nice to have new toys to play with, isn't it?
Minra is one of those things that depends heavily on your application's I/O
pattern. If you have an application that is pretty sequential in I/O access,
then you want read ahead.
From experience, I can tell you that our Oracle Loan servicing application
were beating the heck out of the disks we had. We were only using about 50%
of the disk space but the utilization were very high. We turned on minra=on
and right away cut about 30% disk utilization off. The good thing is that
there were no noticeable cut in application performance.
One other thing you might want to try (not sure how production it really is
in your environment) is no_atime_update. Might help if your email
application does not depend on it. Again, Oracle does not seem to need this.
Derek
-----Original Message-----
From: Steve Losen [mailto:scl@sasha.acc.virginia.edu]
Sent: Wednesday, April 27, 2005 3:03 PM
To: toasters(a)mathworks.com
Subject: Use minra or not?
Please forgive all these questions to the list -- new toy syndrome.
We have deployed a FAS960c pair in a University wide email system with about
30,000 accounts.
We have 3 large Sun Solaris servers running Communigate Pro and they connect
to the filer pair over a FC SAN.
We have mapped 2 LUNs to each server, one from each filer, and striped them
with Veritas. This has nicely balanced the load across both filers. We
built a Veritas vxfs filesystem out of the two striped LUNs on each mail
server and copied the inboxes over. Each mail server holds 1/3 of the
inboxes and operates independently of the other two.
It went into production today and the netapps appear to be more than up to
the task.
We are using mbox format where an email folder is one sequential file.
Consequently our email servers read a LOT more KB/sec than they write, by a
factor of more than ten to one. This afternoon I noticed that our filers
were each reading about 100 MB/sec to 120 MB/sec from the disks. But I also
noticed that this was 25% to 30% more than was being written to the FC SAN.
Each filer has 6G of RAM. Reading disk data at a rate of 100MB/sec should
completely turn over the RAM cache in about a minute, and sure enough, the
cache age was typically 1 minute or less.
I figured that we were reading 25% to 30% more from the disks than we were
writing to the FC SAN because of read ahead. So I set "minra on"
on the volume. Immediately we started reading a little less from the disks
than we were writing to the FC SAN. Also disk util dropped from around 22%
to around 15%. Cache hit percentage dropped a little from 97% to 96%, but
then went back up. (How can cache hits be so high with the RAM cache being
turned over every minute?)
So it appears that in this situation that minra=on is a winner or is there
something I have missed?
Steve Losen scl(a)virginia.edu phone: 434-924-0640
University of Virginia ITC Unix Support