In his blog entry entitled Are SSD-based arrays a bad idea? Robin Harris argues that packing arrays full of disk-form-factor SSD’s is counterproductive. Why? He cites several reasons, including latency, bandwidth, reliability, and cost, but mostly it boils down to squeezing a fast storage media into a slow architecture—much like driving a race car through rush hour traffic or putting wings on a bicycle. Cost and reliability come into play as well, because shoving flash into a disk form factor is less space efficient, less reliable, and more expensive than mounting it on a board.
Enterprise SSD is a young, rapidly evolving market and will continue to evolve until the industry agrees on the perfect SSD architecture and creates standards around it. Expect that to take several years. In the meantime we at Kaminario believe we have come pretty close. We agree with Harris that board-mounted flash makes a lot of sense for reasons of cost, performance, and reliability. That’s why we pack the K2 full of board-mounted PCI flash cards and DRAM. We also hold down cost with our N+1 HA architecture, RAID 10HD data protection (See What You Need to Know About SSD HA and Data Protection and Why Kaminario’s DataProtect is a Big Deal), and the use of industry standard components, the PCIe bus, and market leading Fusion-io technology.
We agree that nothing should get in the way of flash and DRAM’s spectacular low latency. That’s why we designed the SPEAR architecture to use parallelization and other techniques to let SSD show its stuff. SPEAR maintains performance even when an entire storage node fails and it scales performance and bandwidth as it scales storage capacity, so you never have to worry about controller bottlenecks as you do in a classic disk-style RAID box.
And the K2 is the only SSD option on the market today that lets you mix DRAM for superfast writes with more economical flash for fast reads.
The market will continue to evolve and we’ll evolve along with it. In the meantime the question isn’t whether SSD arrays are the problem, but whether SSD should be hobbled by legacy disk architectures. It shouldn’t.
Tags: bandwidth, data protection, disk form factor, DRAM, enterprise SSD, Flash, FusionIO, high availability, K2, Kaminario, latency, PCIe, RAID, reliability, Robin Harris, SPEAR, SSD, SSD array, Storage Mojo
If you toddle over to my site, you’ll see years of SSD promotion, headed with Torvald’s musing that flash will change file systems in 2007. Sun, before Oracle gobbled it, released a “pure” flash storage appliance.
Whether dispensing with file system semantics is sufficient, I don’t think so. The cost difference per byte is never going to close (with flash, Unity’s CMOx and others, who knows?). The real bang for the buck lies in high normal form RDBMS, since such structures are the minimum cover (to bend math jargon) of a datastore. Just being a bit faster processing bloat flatfiles will never cut it.