Flash and cache and flash in cache

Storage Anarchist had a good post on Friday last week discussing the challenges of marketing a new technology like flash SSDs. 

I thought I'd try to contribute a little something to the understanding of how flash is used today in storage systems by contrasting it with how dynamic RAM is used as cache.

Comments

[the storage anarchist]

Thanks for the shout-out.

A couple of small corrections:

First, there are some NAND Flash SSDs that accept writes much faster than disk drives - the ZeusIOPS that EMC is shipping is one example. Indeed NAND is still significantly slower than SDRAM (for both writes AND reads), making Flash less than optimal as a Caching device. There is even data that shows on a $/IOPS basis, DDR2 is about 1/4 the cost of NAND - so if you want to optimize for IOPS, DDR SDRAM is the more practical way to go.

Second point - you don't necessarily have to share ALL of your SDRAM cache across all applications - with Symmetrix at least, you can assign subsets of global memory to different applications (LUNs). These can be dynamically adjusted, and you can even have the system adjust them automatically based upon competing workloads in real time. Admittedly an advanced feature (I'm not aware of any other array that can do this).

Finally, with Symmetrix at least, combining a global memary SDRAM cache with NAND-based storage can result in even better performance than either alone. Mitigating the relative slowness of writes (acknowledge before write) is a key value proposition for SDRAM caches, and this is applicable to Flash SSDs just as it is to spinning rust.

I look forward to continuing the conversation - thanks!