Sizing 3PAR based on EVAPerf results

[slink]

Thank you for the advice. I did meet with HP/presales and we worked through the configurators and came up with a layout. This is all good and I will be going with it. For reference I ended up with 144 SAS 600GB, 24 2TB NL and 16 480GB SSD with estimated usable capacity of ~90TB, realistic IOPS of ~40,000 at ~5ms latency.

So now I have some more wooly questions that betray my ignorance (I've got a few of those on this forum ) . I have to write a requirements document from which will be derived some tests to see if performance requirements are met. Based on the EVAPerf results, which is all I've got to go on really, it looks like most systems that are using the current EVAs are demanding ~9000 IOPS with a mixed workload of 70/30 R/W.

So I was thinking of a requirement that said something like: The primary array shall be capable of delivering a minimum of 9000 IOPS for a simulated mixed workload of random 70/30 Read/Writes using a 4K sample block size over a 1TB RAID5 volume at a latency of <10ms during all system functions including snapshot operations.

Trouble is, I have no real world experience of storage array performance. I don't know if this is actually good, bad or average. Should a 3PAR with the the number and type of disks I've configured be delivering significantly more than this with this kind of test workload or would that be reasonable performance?

[Cleanur]

Assuming Raid 5, with 144x600GB-10K disks @ 70/30 R/W mix with 8K block size, 100% random, zero cache hits you should be capable of delivering in excess of 10K IOps at sub 10ms latency.

If you add in some cache hits say 15% to keep it conservative then you're looking at ~13K IOps so you should easily exceed those numbers before you add SSD into the mix.

Depending on your I/O density with SSD and AO you could see anywhere from 25K to 50K IOps but really just depends on your application mix and the setup.

Now if you want to prove these numbers you have to think carefully about how you go about testing the performance it's essential that any benchmark tool you use is able to reflect a real world workload mix and also that you fully document the test to ensure it's repeatable. I/O Meter is a good place to start, just make sure you setup a decent queue depth to drive the I/O on the 3PAR.

[slink]

Thanks, that's good to know. I was going to document tests either with IOMeter or SQLIO.

[afidel]

If you have a VMWare environment you can use I/O Analyzer (link), basically it will record the real disk transactions of VM's on your existing array and then setup a complete copy of that capture on your new array to play everything back in either a 1:1 or synthetic manner (ie drive the same I/O profile through a single vmdk or replay the capture at an increased pace). It's what I used when I wanted to perf test my 7400 before moving non-production over.