RE: ORION num_disks
- From: "Allen, Brandon" <Brandon.Allen@xxxxxxxxxxx>
- To: "Greg Rahn" <greg@xxxxxxxxxxxxxxxxxx>, "Oracle-L Freelists" <oracle-l@xxxxxxxxxxxxx>
- Date: Fri, 12 Sep 2008 14:04:10 -0700
You could be right, I'm really not sure and have just come to most of my
current conclusions through trial and error. One thing I've noticed, at
least on Linux (OEL4 & 5) is that orion seems to return pretty
consistent results regardless of how high I push the load for a single
execution, e.g., even if I run with num_small 50 (I usually focus more
on IOPS since I work with OLTP systems) and/or num_disks 50, I'll get
about the same throughput as if I run with 5 or 10. I also never see it
spawn multiple processes/threads at the OS level, so it seems to just be
doing AIO from a single process. I've found that I can push the system
much harder if I run multiple orion processes concurrently, so what I'll
usually do is something like this:
1) Create four 4GB files with dd
2) Create for lun files, e.g. test1.lun, test2.lun, test3.lun and
test4.lun, each pointing to 1 of the 4 test files I created
3) Put four orion commands in a script like this to run four orion
commands in the background:
orion -run advanced -matrix point -num_large 0 -num_small 5
-testname mytest1 -num_disks 1 &
orion -run advanced -matrix point -num_large 0 -num_small 5
-testname mytest2 -num_disks 1 &
orion -run advanced -matrix point -num_large 0 -num_small 5
-testname mytest3 -num_disks 1 &
orion -run advanced -matrix point -num_large 0 -num_small 5
-testname mytest4 -num_disks 1 &
4) Run the script
I'll repeat the above test, increasing the number of concurrent
executions until I find the peak performance. Maybe I'm just doing
something wrong with the standard load-setting parameters, but this
seems to be the only way I can get orion to max out my systems.
-----Original Message-----
From: Greg Rahn [mailto:greg@xxxxxxxxxxxxxxxxxx]
I believe num_disks has to do with the number of I/O threads that are
spawned and num_large has to do with the number of outstanding I/Os
that are targeted to be issued.
--
//www.freelists.org/webpage/oracle-l
Other related posts:
