I attended the OpenSolaris Storage Summit in San Francisco. Unsurprisingly, SSDs dominated the proceedings. Sun is touting its hybrid storage pool approach, where SSDs are used to accelerate the ZFS Intent Log (ZIL), a journal, and the L2ARC read cache, to get most of the benefits of SSDs without having to store an entire dataset in expensive SSDs. The presentations on the page are worth having a look into.

Andrew Leventhal commented on how using disk-oriented APIs to access SSDs makes as little sense as using cassette tape adapters to connect an iPod to a car stereo. Sure, it works, sort of, as a short-term compatibility mode, but you are not using the potential of the device. Fusion IO makes SSDs that are connected to the PCIe bus directly and get 5x better performance than the best SATA drive, the Intel X25-E, but it uses proprietary drivers. In one of the video presentations distributed in the conference, Andy Bechtolsheim mentioned he thought a memory channel model is a better fit for the hardware, as the block-device model and all the legacy assumptions behind it is holding back databases from using the full power of SSDs.

A modest suggestion – there is such an API available already: memory-mapped files. In an ironic full circle, an abstraction meant to simulate random access memory on rotating media is actually the best fit to the actual device characteristics. There are also obvious benefits such as DMA.