quick disk performance test ?

[Stuart Abramson]

I wanted to run a quick disk performance test, in response to questions that have come up.
I wrote this little script:
TIME=$(timex dd if=$Disk of=/dev/null count=10000 2>&1 \
| grep real | awk '{print $2}' )
print -n ${DISKA[$I]} "\t" $TIME "\t"
echo 5.12 $TIME | awk '{printf "%.2f\n", $1 / $2 }'
The idea is that I read raw 10000 blocks of block size 512 bytes = 5 MBs and then caclulate the MBytes/sec.
I'm not conducting a rigorous scientific test. I just want to have something to take to the performance group (who will make recommendations back to the UNIX and Storage teams).
Questions:
1. Does this sound reasonable?
2. Is a raw read off the rdsk acceptable for performance measurements? I compared it to a dd if=/dev/vgXX/lvolNN and it came out in the same range.
3. I'm getting values in the range of 2.8 MBytes/sec. Does that sound reasonable?

[harry d brown jr]

What kind of disk drive?

Performance will be based upon speed of rotation (RPM's), IO interface speed, CPU availability, disk caching, ....

live free or die
harry d brown jr

[Simon Hargrave]

This will give you a basic test however

- If you are testing a disk on an array (XP, EVA, EMC etc) then your results will be greatly skewed due to cache.

- You should perform the tests when the system is idle - random access to the disk you are testing again will skew the results.

Your 2.8MB/sec doesn't sound very fast, but it depends again on your server. On an rp54xx I'm getting about 8MB/sec on the internal 36Gb disks, but if you're testing an old A class for example those numbers may be more indicative.

You may be better using a combination of Glance/Measureware to get performance information with "real" data.

[Alzhy]

I will do:

dd if=/dev/vgname/rlvolname of=/dev/null bs=1024k count=nnn

where nnnn should be sufficiently large to overcome the effect of cache.

Note, for LVM.. your RAW device is the lvolname prefixed with "r".

You may also do the same command on raw physical disks.. i.e. /dev/rdsk/cXtYdZ

I would prefer iozone though.

[Stuart Abramson]

They're EMC drives.

[Cem Tugrul]

http://forums1.itrc.hp.com/service/forums/questionanswer.do?threadId=864715

[Cem Tugrul]

hi,
Please have a look my attachment

good Luck,

[Stuart Abramson]

Several people told me that my blocksize for my reads was to small and that fixed me up.

I tried various blocksizes from 512 to 32768 and I/O rate increases from 2 MB/s to 45 MB/s (depending on server, disk, etc.). Block Size of 8192 gives me about 20-30 MB/s on hpux L3000s connected to EMC Symmetrix DMX 2000 with 2 Gb/s FAs:

((BS=8192))
((Count=5120000/BS))
TIME=$(timex dd if=$Disk of=/dev/null count=$Count bs=$BS 2>&1 \
| grep real | awk '{print $2}' )
MBpS=$( FDIV 5.12 $TIME )
printf "%5s %6s %8s %6s %6s %6s %6s \n" \
$LUN $ARRAY $DISK $SP $BS $TIME $MBpS

[Duncan Edmonstone]

Don't re-invent the wheel! There's plenty of free software out there that will do this kind of testing pretty easily - my personal favoutite is iozone:

http://www.iozone.org

But HP also have a bunch more listed here:

http://h20000.www2.hp.com/bizsupport/TechSupport/Document.jsp?objectID=lpg50460#N1053E


HTH

Duncan