Current SATA IOPS info?

[psyches]

I'm trying to reconcile data I'm seeing; a couple of years ago ('05) xmsre posted that one should expect 36 IOPS for a SATA 7200 rpm drive here: thread492-1155465

I remember seeing another where 50 IOPS was quoted, last year (sorry, don't have the thread).

But at

http://www.storagereview.com/comparison.html,

I see IOmeter file system load benchmarks (80% read, fully random, various sizes, varying queue lengths) showing vastly better performance - the very worst performing drive of the lot shows 52 IOPS for IOmeter(1) and goes up from there, and there are many SATA drives (e.g. the WD Raptors) that perform at well over 100 IOPS, and better than some SCSI drives. I also find other recent papers showing this level of performance.

This is a big gap - what gives? Is 36/50 IOPS just old data / really poor drives / not really valid, or is there something fishy with storagereview.com, or what? I don't know of another site that has exact drive model characteristics, and I don't have access to hundreds of different drive types myself. I tend to believe the site because it's more detailed than other info ("36 IOPS for SATA" doesn't account for varying drive types and that obviously would make SOME difference). But I'm not sure.

Thanks for any insights...

 

[comtec17]

The numbers of 50IOPS is the average number most SATA drives perform at. This number also depends on the RPM of the drive and the manufacturuer. For example, a Seagate SATA that EMC uses only performs at an average of 50IOPS but can perform faster by adding more spindles within its RAID type and using lots of cache.

 

[psyches]

Thanks. Do you have the drive model number for the Seagate SATA EMC uses, that's 50 IOPS avg (I assume that's measured for single-sector reads)?

Do you know of a problem with IOMeter or the benchmarking procedures that is showing many, many 7200 RPM SATA drives performing at 80 IOPS or better?

 

[xmsre]

If I use the number 50, it's an average maximum across drive vendors. As IOPS/spindle approaches the maximum, response time increases. At 50 IOPS/spindle, you're looking at response times >50ms. The 36 IOPS number was specifically stated at a 20ms response time. What does your application require?

For some applications (Exchange) anything over 20ms average is considered poor performance (from "Optimizing Storage for Exchange Server 2003"). For SQL, and many other database applications, you'll want an even lower response time, meaning a lower number of IOPS/spindle. Yet for some applications, CIFS file servers come to mind, 50ms or greater may be an acceptable response time. It really comes down to, an IOPS/spindle number has no meaning unless it is associated with a specific response time.

Take a look at the requirements for your application, then consult with your storage vendor on the appropriate number to use.