8 Krypton

throughput and IOPS

in EMC CLARiiON Storage System Fundamentals for Performance and Availability, page 29. it says: This measurement(of IOPS) is independent of request size

so, what's the unit of IOPS. in my opinion, the IOPS is very dependent on the IO size.

Labels (1)
Tags (4)
0 Kudos
3 Replies
8 Krypton

Re: throughput and IOPS

The context of that statement is in actually measuring IOPS...  What they are saying is that the number of I/Os processed per second is a measurement of it's own and is not dependent on the size of the IOs themselves.  ie: 1000 IOPS @ 4KB = 1000 IOPS, 1000 IOPS @ 32KB still = 1000 IOPS.  As IO size varies at a given IOPS, bandwidth will be affected.

Where IO size matters is as size climbs over 16KB and into the 64KB+ territory, the maximum number of IOPS sustainable by a given storage system will decrease.  ie:  IOPS and Bandwidth are inversely proportional.

8 Krypton

Re: throughput and IOPS

Actually it's the other way around: IOps varies as I/O size varies. IOps is a dynamic value, I/O size is a fact. If I/O size goes down, IOps goes up. And between 32kB or 1kB there's absolutely a difference in IOps.

0 Kudos
8 Krypton

Re: throughput and IOPS

Knowing the author, I can confirm what his intent was. But I think he caused some confusion.

The measurement of IOPS is IO size agnostic. E.g., If I am doing 1000 IOPS at 8 KB, or 1000 IOPS at 32 KB, I'm still doing 1000 IOPS. A pound of lead weighs the same as a pound of feathers.

The performance of a system when using IOPS as a measurement is dependent on the IO size. So, if my machine has the capability of performing 1000 IOPS at 8 KB, it is unlikely to have the capability of doing 1000 IOPS at 32 KB. My car can do 150 Kilometers per hour but only 90 Miles per hour.

The conversion from MB/s to IOPS is dependent on IO size. This is a linear relationship. 1000 IOPS at 8 KB is (1000 * 8 )/ 1024 MB/s = 7.8 MB/s

However 1000 IOPS at 1 MB IO size is 1000 * 1 = 1000 MB/s