Nomad_Soul - Fotolia
- Arun Taneja, Taneja Group
All-flash storage arrays are hot. Suddenly, there are dozens of available systems from startups and established vendors alike. And a lot of them claim they can deliver all-flash arrays at prices lower than high-end 15K rpm hard disk drive (HDD)-based arrays. Not to be outdone, at least one "hybrid" array supplier claims it can deliver all-flash array-like performance at well below high-end HDD array prices. All of this makes selecting the right system for your environment that much more complicated.
Purchase criteria for all-flash arrays must begin with performance, and there are three important factors to consider: IOPS, latency and throughput. Depending on your application, either IOPS or throughput will come into play. IOPS will reign supreme for small-block I/O-bound applications, while throughput is key for large-block I/O-bound applications. Latency is important in all cases. You'll quickly discover that existing HDD-based designs that have been modified to become all-flash arrays don't deliver sizzling performance. Flash is very different from hard drive media; it performs differently, fails differently and is managed differently. Modifying existing controllers only goes so far. To extract every bit of performance from flash requires new thinking that generally comes from starting the design from scratch. I have seen one exception: the Hewlett-Packard 3PAR StoreServ 7450 all-flash array performs like a specially designed all-flash array because of the custom ASIC inside the 3PAR system.
Despite the "software-defined everything" mantra so prevalent nowadays, all-flash arrays are hardware-defined today. They're proprietary and come in complete packages from each vendor. Most offerings are built around solid-state drives (SSDs) as the basic element, but some start with flash chips. SSD-based systems rely on a lot of functionality coming from the SSD itself, like garbage collection and wear leveling. You get whatever comes from the SSD vendor. However, designs based on flash chips can perform such functions at the array level rather than at SSD level, which often results in added performance, resiliency, better density and lower power requirements. Those benefits are reflected in the price of the system. IBM, Skyera and Violin fall into this category.
Inline data deduplication and compression remain controversial topics in the all-flash storage array world. With some arrays, those functions are built in and you can't disable them. Equally important, the pricing reflects the assumption of a 4:1, 5:1 or 6:1 reduction factor to arrive at the magical "equal to HDD prices" value proposition. Other arrays lack these features entirely. There are two things to note here. First, deduplication is awesome if it applies to your application. Virtual desktop infrastructure and virtual servers are at the top of that list, but it doesn't do much for databases or video streams, among other applications. The applications you'll run on an all-flash array will tell you if the $/GB touted by the vendor is applicable. Second, deduplication and compression are excellent features if they don't sap performance, which is the reason you're buying an all-flash array in the first place. So you need to view these features within the context of performance. Basically, this means the vendor must counterbalance performance-sapping functions with added hardware that neutralizes the impact. Ideally, you should be able to engage or disengage these functions at a LUN level.
The next consideration is performance consistency. A well-designed system must perform consistently, with its latency staying within a narrow band. You'll likely expect an all-flash array to support a multitude of applications, using a variety of block sizes, all requiring excellent performance. Ask the vendor to show you how the array performs at each block size, rather than just the average of all these block sizes.
Data protection is a different animal when applied to flash. We're so used to using RAID as the data protection mechanism in the HDD world that our first tendency is to bring it over, blindly, to the flash world. Flash is a random media, with no seek times or rotational latency issues to worry about, while RAID was designed with those issues in mind. RAID concepts can be brought over to flash, but you'll have to rethink data protection from scratch. For instance, there's no need to gang up only five or six drives in a RAID 5 or RAID 6 set as much longer stripes can be used. The rebuilds can be done differently since flash typically fails differently. Check out how data is protected and what happens when a flash unit fails. Expect much better data protection than you get with HDDs.
Any all-flash array will also offer considerable advantages over HDD-based arrays in terms of power, cooling and density. But you should compare all-flash arrays with each other, as these systems can differ greatly in those dimensions. Power and cooling requirements generally follow density. The industry is still learning how to get the maximum bang for its buck in all-flash arrays, and some vendors do a much better job than others.
Any discussion of all-flash arrays would be incomplete without mentioning scalability. There are a variety of products that are scale-out while others are scale-up. As you evaluate all-flash storage arrays, make sure the starting point gives you sufficient performance and capacity for today, with enough room to scale over time in both dimensions.
A few years ago it was OK to buy an all-flash array that came without all the enterprise-caliber storage applications and services, such as thin provisioning, snapshots, replication and cloning. Today, I wouldn't touch an all-flash array without a full complement of those features. And the same applies to ease of use and manageability.
Not every all-flash array is alike, and what you purchase depends on the type of applications you want to deploy with them. Any all-flash array will give you better performance than an HDD-based array, but that would be like buying a Porsche without checking out a Lamborghini and a Maserati first.
About the author:
Arun Taneja is founder and president at Taneja Group, an analyst and consulting group focused on storage and storage-centric server technologies.
all-flash array vendors need to offer advanced enterprise features