Iron out the kinks and all-flash servers just might have a future

Comment Nutanix's NX-9000 is an all-flash, scale-out, hyperconverged server/storage/networking system, with a single logical pool of SAN storage aggregated from the NX nodes in a cluster. In other words, it's a clustered, disk-less server.

It's disk-less, and so less prone to mechanical and electro-mechanical errors and needs less power to run. It is also fabulously faster at random data accesses than a disk-based system.

Two questions:

Why, on this basis, would anybody want a server with direct-attached disks?

What role, if any, does a networked SAN have when you use all-flash servers?

The NX-9000 has limited capacity; 9.6TB using 1.6TB SSDs. Six terabyte disks are appearing and soon enough we should see 8TB and 10TB versions in 3.5-inch form factors, and 2TB and 4TB ones in the 2.5-inch form factor. Disks are cheaper per terabyte than SSDs, so if you don't need and can't afford all-flash performance a disk-based server will be cheaper and hold more data...except that a flash server can deduplicate and compress data inline and so increase its effective capacity.

Also, SSDs, as well as PCIe flash, are increasing their capacity, with 3D NAND technologies promising to double or triple SSD capacities. This time next year we could be looking at an NX-9000 with 6x3.2TB SSDs, near enough 20TB which, with 2:1 or 3:1 data reduction, could mean 40-60TB effective capacity.

Disks are an ineffective method of storing random data and getting it fast to users. Flash is far, far faster. Moreover, fast data access in our mobile, need-it-yesterday world is prized more than marginal increases in cost.

We could be seeing the first concrete sign of a move to all-flash servers for performance-sensitive use-cases.

In which case the second question is pertinent: why, when you have all-flash servers with server-SAN capabilities, do you need a networked storage array?

Good question, but the answer lies in capacity.

A flash server can run an entire working set of data in DRAM and flash, if the working set fits in the available space. However, working sets are subsets of complete data sets and they run up into the petabytes and tens if petabytes range, and beyond that with the largest data sets. There isn't enough flash in the world to hold this data so it has to go on disk arrays, which then feed it to servers as and when required. (Although 3D TLC flash might change that view.)

Secondly, a hyper-converged system is a single box. You need more performance and networking bandwidth and storage capacity? Great. Get another box.

If you need more of just one of them separately – more capacity say – then tough, you can't have that.

OK. Let's have performance nodes, and networking nodes, and storage nodes. You just lost out on hyperconvergence simplicity and we already know what a storage node in a cluster is called; it's a networked array component.

Storage arrays need to evolve in a hyper-converged, clustered, all-flash server world and become capable of joining the server-side SAN as a transparent and seamless, albeit slow, partner. Secondly, they need to adopt faster networking to reduce network latency. Thirdly, they need to be able to stream server-focussed working set data into servers, avoiding or mitigating random IO wherever possible. And lastly, they need to be the backing or nearline tier in a multi-tier, clustered server, server-side and networked SAN construct, with automated data movement.

Is this feasible? It will need a SAN vendor with great connections to the server-SAN world. Who might this be? Which SAN vendors have the right components on which to build such products?

• Cisco — not likely, as it has only a nascent SAN business with Invicta flash arrays.
• Dell — could be, as it has both servers and SANs in its product armoury
• EMC — yes, it has both server-SAN and networked arrays in its locker
• HDS — in theory yes but it has no server-SAN capability yet
• HP — yes, it has servers, StoreVirtual and 3PAR SANs
• IBM — servers and SANS but no server-SAN capability - yet
• Oracle — yes, think engineered systems and the FS1 arrays
• NetApp — no servers but an ever-willing collaborator

The flash and server combination could yet upset a few SAN apple carts. ®