Netapp and Equallogic experiences

robocat - Are you affiliated with Netapp?

I understand I need to simulate what I intend to do in my production environment to see true results.  My issue is HOW exactly do I do the testing to produce accurate results.  I know of using iometer to simulate disk activity, and jet stress for Exchange, but aside from that I'm stuck.  There must be more to it, because that seems too simple.  I've never really been faced with having to do performance testing, as management always purchases the standard DL380 with our current environment and we've not had performance issues.

>Are you affiliated with Netapp?

No, we just use Netapp on our corporate network in a mixed unix/windows/vmware environment. Feel free to confirm what I'm saying by testing this yourself.

>My issue is HOW exactly do I do the testing to produce accurate results.

First you need to know the characteristics of your I/O, e.g. how many I/O sec does each app use, how large is the average I/O size, etc. You can collect these stats using standard performance tools.

You will probably notice that for the apps you mention (SQL, Exchange, ...) that you get mostly random I/O, small blocks (4K-16K-32K) and that the IOs/sec are relatively small.

Use this info to start performance tests with your favorite benchmark tool. E.g. set it to test for random IO, 8K block size and 4 or 8 simultaneous outstanding IO threads. Do this from more than 1 VM at the same time.

Important is to use a test file that is relatively large (at least 4GB) to avoid that all blocks are cached in RAM after a while.

For the results, don't look at throughput but look at latency(response time) and IO/sec.

My coworker has been running some tests on the Netapp.  So far he's only done it with small files (2 megs, I think), and running 8 workers simultaneously, 4 per VM (one ubuntu, the other debian).  Tested using NFS vs iSCSI, and NFS had better IOPS than iSCSI, which doesn't surprise me in the least.  What I am confused by is the technician from the vendor I've been dealing with says that architecturally it should preform better on random tests than sequential.  We're producing the opposite results, so now he's confused.  Enabling jumbo frames and flow control improve performance (of course), but it's proportional to running it without jumbo frames.  Do you see this this in your environment?  While sequential would be a-okay for our home directories, the random portion of it concerns me with databases....especially because my Exchange server is heavily read-biased.

In regards to your previous question re: improvements to the file system, I was referring to the fragmentation issues that can occur if the filer get full, and people have to "defrag" and run the scan reallocate, or something like that.  This apparently allows them to handle more full filers better.

How long have you been using Netapp?

>it should preform better on random tests than sequential.  We're producing the opposite results.

What do you mean by "better" ? To what are you comparing ?

>This apparently allows them to handle more full filers better.

It is still best not to fill it up to capacity.

We've been using NetApp since 2002.

Sorry so slow....very hectic week.

Random workers get poor performance vs. Sequential works when you compare the total mbps and iops of workers.  The random workers numbers are one tenth of what the sequential are.  Doesn't matter if we use NFS or iSCSI, or the load of read vs. write.  Random are all approx 10-15% of what Sequential performs.

Technician said that unlike a typical SAN architecture, the random workers will perform better on a Netapp filer.  I don't understand how they can, but that's aside from the fact.  I've been questioning as to why my results do not reflect what I'm being told, but have not been getting anywhere or any good answers.  As a result of this, I'm unsure as to whether or not we're testing properly, or if the technician is just full of BS.