There was a big flash, and then the dinosaurs died
This is how it was supposed to end. The legacy SAN and NAS vendors finally realize that Flash is fundamentally different from HDDs. Even after a decade of efforts to completely assimilate Flash into the legacy architectures of the SAN/NAS era, it’s now clear that new architectures are required to support Flash arrays. The excitement around all-flash arrays is a testament to how different Flash is from HDDs, and its ultimate importance to datacenters.
Consider what happened in the datacenter two decades ago: HDDs were moved out of networked computers, and SAN and NAS were born. What is more interesting, however, is what was not relocated.
Although it was feasible to move DRAM out with technology similar to RDMA, it did not make sense. Why move a low latency, high throughput component across a networking fabric, which would inevitably become a bottleneck?
Today Flash is forcing datacenter architects to revisit this same decision. Fast near-DRAM-speed storage is a reality today. SAN and NAS vendors have attempted to provide that same goodness in the legacy architectures, but have failed. The last ditch effort is to create special-purpose architectures that bundle flash into arrays, and connect it to a bunch of servers. If that is really a good idea, then why don’t we also pool DRAM in that fashion and share with all servers? This last stand will be a very short lived one. What is becoming increasingly apparent is that Flash belongs on the server – just like DRAM.
For example, consider a single Fusion-IO flash card that writes at 2.5GB/s throughput and supports 1,100,000 IOPS with just 15 microsec latency (http://www.fusionio.com/products/iodrive2-duo/). You can realize these speeds by attaching the card to your server and throwing your workload at it. If you put 10 of these cards in a 2U-3U storage controller, should you expect 25GB/s streaming writes, and 11 million IOPS at sub millisecond latencies. To my knowledge no storage controller can do that today, and for good reasons.
Networked storage has the overhead of networking protocols. Protocols like NFS and iSCSI are not designed for massive parallelism, and end up creating bottlenecks that make crossing a few million IOPS on a single datastore an extremely hard computer science problem. Further, if an all-flash array is servicing ten servers, then the networking prowess of the all-flash array should be 10X of that of each server, or else we end up artificially limiting the bandwidth that each server can get based on how the storage array is shared.
No networking technology, whether it be Infiniband, Ethernet, or fibre channel can beat the price and performance of locally-attached PCIe, or even that of a locally-attached SATA controller. Placing flash devices that operate at almost DRAM speeds outside of the server requires unnecessary investment in high-end networking. Eventually, as flash becomes faster, the cost of a speed-matched network will become unbearable, and the datacenter will gravitate towards locally-attached flash – both for technological reasons, as well as for sustainable economics.
The right way to utilize flash is to treat it as one would treat DRAM — place it on the server where it belongs. The charts below illustrate the dramatic speed up from server-attached flash.
The servers are beckoning the long-lost storage child to come back home. The all-flash arrays mark the beginning of the migration of flash out of the stronghold of the legacy SAN/NAS architectures. Next stop is home.