Sign in with
Sign up | Sign in

How Numbers Disagree

Enterprise SSDs: How to Prove Real World Performance
By
Brought to you by What's this

To recycle our shoveling metaphor from the last article, one could say, “My back is shot! I just moved 50 wheelbarrows of stuff from the driveway to the back yard.” As a listener to this remark, you don’t know whether to be sympathetic or laugh. Did the speaker move 50 loads of dense, heavy gravel or light, puffy insulation scraps? The load being moved makes all the difference.

This is why you can see a drive achieve 25,000 IOPS in Iometer and still not know much about its true performance. The same drive might only realize 15,000 IOPS in another benchmark. The data being cycled through the drive, the patterns in which it writes to media, and other factors comprising the workload make a significant impact on the final result.

“It’s not correct to say it’s 14,000 IOPS,” says Seagate applications engineer Jeff Nowitzke. “People should say it’s 14,000 SPC-1C IOPS. The differentiation needs to be there, because when the Storage Performance Council runs [its SPC tests], it consists of a complex workload characterized by various transfer sizes and transfer modes. Some are sequential, some are random, some are read, and some are write. It’s not the same as 100% 4K read or 100% 4K write IOPS. It’s a separate unit of its own that is used primarily for comparison purposes.”

Everyday data center apps will write with I/O block sizes ranging from 2K to 256K. Highly sequential loads, such as backup and media files, can have blocks of several megabytes. As any thorough benchmarking report will show, performance will vary considerably across this spectrum. Even random data at large block sizes will approach sequential write speeds because so many NAND pages are being written in consecutive order to accommodate those larger block sizes.

When an organization fails to standardize its testing on preset scripts, the problems only magnify, because different technicians are prone to using benchmarking apps with different parameters. Clearly, there needs to be one go-to tool able to use a credible, consistent workload with predictable parameters so all evaluators can assess drives with a common base of metrics. It’s not easy or cheap, but such a resource does exist.