When new products release, I look for elements that stir me. In the tech realm, components perpetually grow bigger and faster, but without innovation, that can become a liability. What transforms upgrades from mere change to worthy innovation is intelligence. Is the new product smarter than it was in the last revision/model/version?
The HP 3PAR StoreServ 20800 series easily checks the boxes for bigger and faster, but the aspects that jump out as smarter are Asynchronous Streaming Replication and the near-atomic scalability, particularly focusing on the minimal instantiation size. I’ll explain what I mean on this latter point farther down.
Flash has changed the physics of replication, but I honestly hadn’t realized it until the HP briefing on the new 3PAR 20800. Prior to flash, synchronous replication was a factor of inter-array network latency and was restricted in implementation accordingly. A typical high-bandwidth, 1ms link was all an organization required. Then came flash and mere microseconds of latency (200-500 on average).
Imposing synchronous replication on an all-flash array would destroy most of the rationale for purchasing it in the first place. Minimum response times would rise from 200μs to 1-2ms at a cost of 5-10x performance degradation. You might as well bring back the spinning disks.
3PAR has an answer for that in its next 3.2 OS code and it’s called “Asynchronous Streaming” replication mode. In streaming mode, 3PAR writes delta sets in cache that are quickly sent to partner arrays with RPOs as low as 1ms and crash consistency throughout.
What happens if the replication link slows down or a write burst occurs, though? Great question. While cache remains, the deltas build up but continue sending. Eventually though, this condition can lead to a full cache. At that point, 3PAR suspends random replication groups (RGs) and takes volume snapshots from which it restarts replication. It effectively reverts to “Asynchronous Periodic” replication mode. In future releases, QoS will be added to the feature set to allow RGs to be prioritized. Through all of this, host latency continues flying unimpeded.
The 3PAR StoreServ 20800 is also smarter when it comes to scalability, though I argue it differently than HP Marketing does. In fact, I think the graphic above is poor or even incorrect positioning. Value, scale and performance all matter to me, and I argue that the 20800 is HP’s answer to all three.
The 3PAR 20850 starts incredibly small with just two nodes and four disks, which I hear can find a way in the door for as little as $75K. I doubt anyone would quite do this, but this means an organization could feasibly get a 3PAR AFA on the floor with ~14TB of usable storage (with 3+1 RAID-5, plus any deduplication gains) for less than $100K and incrementally add a new 3.84TB disk each month to scale up. Then, of course, it scales to 3.9PB raw (non-dedupe usable = still a lot), which is an amazing span from such a humble start.
That scenario is obviously hypothetical and foregoes the wisdom of forward planning (if under-scaling and constantly restriping), but it emphasizes the flexibility of the product line. I’ll be very interested to see customer feedback when arrays hit the floor. Is there a minimum (or maximum) configuration where the array hits its sweet spot? Other AFAs have those, so I don’t think that’s a downside, just an architecture choice. It comes down to capacity, IOPS, bandwidth, and even interconnect/backplane (especially when we’re talking about huge 4TB disks). Marketing (everyone’s) will claim domination on all, but customers will discover the truth.
Wrapping up, the HP 3PAR StoreServ 20800 series is a great advancement for a proven, performing product line, and it’s definitely a worthy contender in the incredibly competitive flash storage space.