Friday, 28 October 2011

Time to Store All Cloud Data in Dynamic Memory

It is fairly clear we are at a point when purchasing personal workstations consisting of tablets, ultrabooks and laptops that only solid state drives, SSDs, make sense. There are huge performance benefits and the generally lower capacity of an SSD compared to a hard disk drive, HDD, of similar cost is not a problem with apps and data rapidly migrating to the cloud.

One might imagine that the cloud data itself, though, being so massive, must necessarily be stored on massive arrays of HDDs that is currently the case. It therefore stretches belief to find recent research indicating that the most cost effective way to store cloud data from now on is to use massive arrays of standard dynamic random access memory, DRAM. The article by Scott Fulton entitled ‘Stanford: Move the Cloud from Disk to DRAM’ summarises the paper by Ousterhout et al that introduces the punchy concept of RAMClouds. The argument relies on the 3-year cost of ownership for large datasets taken from the second figure in the paper:

2011-10-28 SNAG-00

I expect this to be an historic turning point for cloud infrastructure and totally unexpected. RAMClouds will have an absolutely major impact on cloud services from the performance perspective and well as the design and siting of data centers. I can’t wait to see the effects.