va 7100 slow throughput

[Charles McCary]

Group,
Hi - I have a new va7100 that's only showing around 35-50mb per second throughput. I used the following commands to verify this:

time dd if=/dev/rdsk/cXXXXX of=/dev/null 1024k count=50

This will read 50 MB off the disk device (you can set this to whatever value you want but 50 is sufficient for a good read test) and give you a time in seconds to do it.
eg.
50+0 records in
50+0 records out

Then divide 50 by whatever your time is to get mb / per second.

any thoughts on why this is running at below half of the advertised rate - things looked to be cabled correctly.

tx,
C

[A. Clay Stephenson]

Hi Charles:

Off the top of my head the maximum sustained rate is ~ 90MB/s. You don't mention what your host is, but it's rather difficult to push much beyond your observed rate through a single i/o channel. Glance will probably give you a better (more reliable) performance metric. I would also do your test with more than one server (or at least more than 1 dd).

You can save yourself a bit of clock watching by preceeding your dd command with 'timex'.

You might also want to create a /dev/zero device node (mknod /dev/zero c 3 0x000003) to serve as an input source to measure write rates.

Clay

[Enes Dizdarevic]

There is another discussion on va7100 throughput in this forum. You can follow this link to see it:
http://forums.itrc.hp.com/cm/QuestionAnswer/1,1150,0xe85e46ff9277d511abcd0090277a778c,00.html
I recommend using in this discussion mentioned tool iozone to check performance.

[A. Clay Stephenson]

Hi Charles,

One more thought. At the very least your test is distorted by the time required to load dd and prepare it for execution. This overhead can be adjusted for by doing this command:
timex dd if=/dev/zero of=/dev/null bs=1024k count=50. This value can then be subtracted from your original dd command to give a somewhat better figure for the i/o rate.

[Vincent Fleming]

dd is a notoriously bad performance tester. Because it does I/Os one at a time - waiting until the first returns before issuing the next - it defeats all paralellisms in the O/S. For example, it defeats Tagged Command Queueing. TCQ allows the O/S to queue several I/O's to the device at a time (up to 255 commands). This is important with RAID units such as the VA7100 because it has many disks and can issue I/O's to multiple disks at the same time - and respond to multiple I/Os from the host simultaneously.

To help mitigate this behaviour, try using a different benchmark, or at least execute multiple dd's at the same time - 5 to 10 of them.

Good luck - I'm sure you'll see the 90MB/s if you do two things: use multiple dd's as I suggest, and use both controllers in the array (do the dd's over both channels).