Some flash array vendors want to make the fastest flash array they can. Pure’s array isn’t the fastest, that is by design. Pure Storage want to bring the benefits of flash storage by being price competitive with hard disk based arrays. Pure is targeting being a mainstream storage array that uses flash. Pure want to replace the spinning disk in your datacentre. When I say Pure isn’t the fastest flash array I mean a single array won’t do a million IOPS, only a few hundred thousand. To put this in perspective an all hard disk array without cache would need eight racks of disks to hit 200,000 IOPS. Not the fastest all flash array is still a very fast array. The purchase price for usable capacity may be similar to hard disks. The price per unit of performance is much lower than hard disk.
Flash is much more expensive per TB of raw capacity compared to hard disks. Pure needs to use a limited amount of flash and optimize capacity. They use deduplication and two types of compression to maximise the usable capacity of the flash. Pure’s figures suggest that the usable capacity is about five to ten times the native capacity of the array. Different workloads have different capacity savings depending on their commonality of data and compressibility. A consequence of this data reduction is that your Pure array doesn’t have a fixed amount of usable capacity. Because some compression is done as a background task an array with low disk space one day can have more free space the next morning. On the flip side migrating workloads off the Pure array won’t always increase free capacity. Just like any other deduped array only unique blocks removed result in free space.
The Pure array looks like a conventional array with two controllers and some shelves of storage. Since Pure has an all flash architecture the storage shelves contain SSDs, lots of SSDs. The controllers are x86 servers and use SAS and Infiniband to connect to each other and the storage shelves. The array’s CPUs are always making the data reduction better. More compression and pattern matching keeps controller CPUs at 100% utilization. This is a bit of a surprise when you come from other arrays that get concerned when the controller CPU is above 70% utilization.
Some of the really interesting parts are the non-technical parts of the Pure story. They have a Forever Flash program that keeps your maintenance costs on the array reasonable over years. Even better if you keep buying maintenance in the right way you get no (additional) cost upgrades to a new array every three years. Yes, you read that right, buying maintenance can cover hardware upgrades. The arrays are even designed for incremental upgrades. The controllers are stateless so can easily be swapped for newer and higher performance ones. An interesting part of that is that the entire array is capped at the performance of a single controller. If one controller is out of operation there is no performance penalty. Controller replacement or firmware updates are at no performance impact, potentially an in-hours activity. In hours upgrades without performance sacrifice is important for 24×7 business. The other fun part of Pure is that their arrays phone home with performance information. Having real customer systems reporting the effectiveness of dedupe and compression is powerful.
Who is going to buy Pure arrays? Customers who are currently replacing a disk based array with performance problems. Customers who are disappointed by flash based tiering and caching solutions that don’t match their workloads. I also think that the move to an annual fee based ownership will be very appealing to a lot of customers. The hold back is that there is a lot of smart software here, what if it’s not so reliable? This is where the phone home feature gives confidence by reporting the actual effectiveness of the smart stuff.
The Pure story is very interesting. The Nerf war that erupted half way through their presentation at Virtualization Field Day 3 was a bit of fun to break up the large amount of tech. By the way the picture at the top is of a Lego version of the Pure array, the All Lego Array.
Disclosure: I attended Virtualization Field Day 3 as a guest. Nobody at VFD3 nor the presenters has paid for this post or had any editorial input. Any omissions and inaccuracies are mine.
© 2014, Alastair. All rights reserved.