• 06/27/2014
    7:00 AM
  • Rating: 
    0 votes
    Vote up!
    Vote down!

The Evolution Of The All-Flash Array

Vendors are adding data deduplication and other features to their all-flash arrays, making the storage systems more than just speed demons.

Recent product releases from several leaders in the all-flash array market, including HP and Violin Memory, are further indications that we’ve reached the point where all-flash arrays are more than just the storage equivalent of a top fuel dragster. As flash transitions from being a point solution for the most performance-critical applications to more general use cases, all-flash arrays are adding data services and data efficiency features that at least equal those of their disk-based brethren.

Violin Memory has long been a leader at the "go fast" end of the all-flash array market, but its products have lacked even basic data services. Its model 6000 all-flash array could do a million IOPS, and, thanks to some unique features of Violin’s VIMM flash modules, deliver those IOPS with consistently low latency. However, if you wanted to take a snapshot, you’d better use the snapshot provider in your hypervisor or volume manager.

Part of Violin’s problem, of course, was that its management, from its founding through the ouster of CEO Don Basille, came from the semiconductor, not the storage, part of the market. Naming the company Violin Memory, not Violin Storage or Violin Systems, was evidence of this semiconductor-sourced management.  These managers just didn’t understand that storage folks need data services and want a system with hot swappable modules in the front, not circuit boards that plug in the top.

Violin’s new Concerto 7000 and its Windows Flash Array -- the latter which could be a killer SMB3 server for Hyper-V hosting -- demonstrate that Violin’s new management better understands the storage market. The 7000 is based on Falconstor’s storage virtualization engine, which Violin has licensed and essentially forked for its applications. A pair of virtualization engines in the 7000 provides data services from basic snapshots, metro-clustering, CDP, and scalability controlling up to 280 TB of 6000 AFAs in an intelligent-shelf model where each shelf handles the basics of data protection.

Where Violin added data services to its all-flash array, HP is adding data reduction -- in particular data deduplication -- to the already feature-rich 3PAR 7450 all-flash array. HP claims that, after deduplication, a 7450 can cost less than $2 per gigabyte, which is well below the cost of providing that storage with 10K or 15K RPM disks. HP is starting slowly with deduplication only, no compression, but promising compression in a future firmware release that can be applied in the field.

At this point, let me say that I think HP made a serious mistake naming the 7450. I understand it’s based on the four-controller 3PAR 7400 platform, but I still think the fastest model in a product line should have the highest model number. Just the fact that 3PAR has a 10,000 disk-based model makes the 7450 seem less impressive than it should.

While startups like Pure Storage, SolidFire, and XtremIO (acquired by EMC) have built all-flash arrays with data deduplication, the upcoming software update will make the 3PAR 7450 the first storage system designed by a major vendor with inline deduplication. In the past, when data reduction was added to storage systems like EMC’s VNX and NetApp’s FAS, it either ran as a post-process, came wrapped in all sorts of warnings about how it might negatively affect performance, or both.

As I’ve written previously, flash and deduplication are two great tastes that taste great together. Since flash handles random I/O as well as sequential, there’s little to no performance loss. Add in that a bigger CPU to perform data reduction costs significantly less than the amount of flash it can save, and, as far as I’m concerned, data reduction, along with data services, is becoming table stakes in the all-flash array market.

While HP had the resources to adapt its deduplication technology from the StoreOnce backup appliance to 3PAR, other vendors will be able to get data reduction into their products faster by licensing the technology from Permabit, as Violin leveraged FalconStor’s tech for data services. While few of Permabit's licensees like to make the fact they bought, rather than developed, their tech, Permabit claims to have four all-flash array vendors among its 14 licensees.

All-flash arrays are becoming mainstream, especially in larger datacenters that can isolate their lower performance applications to disk-based systems. As a result, you should stop looking for as much performance as possible and start insisting on data services and efficiency just as you do for disk systems.



Great article Howard. The all flash array is key as companies adopt the cloud. Speed is such a huge factor. It will be interesting to see what vendor emerges from this.

Storage Services

Howard, overall I like the idea in your article and have a few thougths that I think are worth discussing.  Full disclosure, I am with the IBM though these are my personal views. 

First, I think there is a place for deduplication with flash.  I tend to think the one application that really benefits without suffering is VDI.  This is a fairly narrow segment of the market and these features are just as easily if not more cost effectively achieved with application software such as Atlantis ILIO.

Second, there are many places where dedupe and flash do not go together well.  An excellent example is databases.  It is difficult to achieve data reduction with deduplication for databases (outside of the narrow slice of ground some attempt to occupy with the idea that dedupe helps if you put dev, test and production on a single array).  In real world testing, real time compression is more effective with databases.  This is a significant point because most of the applications that use flash today are databases.

Third, the suggestion that  deduplication can be done without much impact performance is, in my experience, incorrect.  Compare an all flash array without dedupe with one that is not implementing it and I promise you will see significant latency impact on the order of hundreds of microseconds.  This is significant because the reason most people benefit from flash is because of the low latency.