Tiered networked storage array startup Nimble Storage won't be struck by the coming EMC Lightning server flash product because its customers don't need it.
The Nimble array uses multiple tiers of storage, including NVRAM, flash and disk, to provide large capacity and avoid disk latency when responding to data IO requests, as well as having the essential feature of sharing storage resources between servers. Is its architecture threatened by flash in servers, PCIe-connected flash storage?
Talking about this question, Nimble's CEO, Suresh Vasudevan, said there were three questions to keep in mind:
- Is the flash a cache or a storage tier?
- If it is a cache does it do read caching?
- If it is a cache does it do write caching?
In his view, treating the server's PCIe flash as a storage tier – the Fusion-io approach – means that it stores an application's entire working set in the server and takes the primary data storage role away from the storage array: "It's only relevant for data not going to the back-end array." You get a very fast response to IO requests but the cost per GB is high and the flash and its data contents is not sharable between servers.
Vasudevan says: "The number of applications where latency in microseconds matter is very small. We are not going after such applications. We focus on mainstream applications where cost/performance and cost/capacity are more important."
If the flash is a server's cache and does read caching then that is great for many applications but not so good for applications that need fast write IOs as well as read IOs. The cache does write caching that solves a write latency problem but adds complexity and may affect virtualised app mobility as vMotion of the app could be prevented or hindered by the cached writes.
His company's array provides write caching through its serialised layout and use of solid state storage and this is faster than a disk drive array, more cost-effective than all-flash arrays and server flash tiers, and also server flash caches. Mainstream applications need to execute faster by freeing themselves from disk drive array hobbles but not at the expense of or with the limitations of server flash tiering or caching. In logistics terms customers in NImble's market need a FedEx truck, not a customised Ferrari.
What's VMware going to do?
Vasudevan thinks it interesting that VMware is not involved in managing server flash memory and asks whether it's likely that in two or three years it will take on that task, much as it has been taking on more of the job of managing access to networked storage. He asks rhetorically why wouldn't it do that; "VMware must see the logic of a cache managed by the hypervisor. VMware understands the block layout involved better than anyone else." ESXi could treat the cache as a tier or a cache. Ditto Microsoft and Windows 8/Hyper-V.
How is Nimble doing? "At our stage of growth every quarter had to be a record quarter. … We now have 22 sales territories … and around 150 people; it was 38 in December 2010." Customer adoption momentum is still growing and it should soon reach the 300 customer point: "a benchmark this quarter hopefully [and] a very big deal for us."
Nimble is sticking to its knitting and continuing to grow fast. It is knocking at an open door as the cost and complexity and slowness of disk drive arrays is being questioned more and more, it says. Customers need a simpler, cleaner shared storage architecture, one unencumbered by by legacy storage designs – and Nimble Storage is hoping to fill that gap. ®