Wikia’s Artur Bergman recently gave a talk at Velocity about SSD adoption that has generated a lot of buzz.
The video can be viewed here.
Warning: the video is rated PG-13 for language and adult situations.
The focus of his talk was that the relevant metric for data center storage is $/GB/IOPS. He showed how adopting SSDs provides orders of magnitude more IOPS over traditional drives and thus will prove themselves more economical.
In addition to greatly increased read and write speeds, the switch to SSDs helps to relieve the bottleneck of random access seek times. Random I/O is often at the core of performance and scalability problems. But as a result of the performance improvements that SSDs bring, he argues that we can once again use standard practices such as monolithic database and file servers and that distributed storage systems are thus irrelevant.
There’s little argument that SSDs will continue to gain a foothold in the data center. But should you switch everything to SSDs? What are some cases where $/GB/IOPS is not the only relevant metric?
Everyone is focused on the explosion of big data, but not every application is I/O-based. Video encoding or scientific research applications often benefit from beefier processing capabilities than I/O speeds.
No matter how fast your database server is with SSDs, if it gets knocked off line, it really doesn’t matter. Distributed data storage systems such as Hadoop and Cassandra are also designed for availability, keeping multiple replicas of data so that if some nodes die, data and services are still available.
Not All Data is Equal
The process of moving ‘cool’ data to an archived location is still an observed practice. Companies may not need blazing speeds for all data and so blindly optimizing for IOPS over raw capacity isn’t always the right way to go.
Don’t get me wrong — I love the SSDs in my laptop and workstation. And over time–in the right cases–SSDs will provide another means in our ongoing effort to squeeze more performance out of our systems.