Disclaimer: I recently attended Storage Field Day 10. My flights, accommodation and other expenses were paid for by Tech Field Day. There is no requirement for me to blog about any of the content presented and I am not compensated in any way for my time at the event. Some materials presented were discussed under NDA and don’t form part of my blog posts, but could influence future discussions.
In talking with people about Pure Storage, some of the feedback I received was that they were a “one-trick pony“. By that I mean that people thought all they did was offer an all-flash array and nothing more. I think there’s always been a lot more to Pure than just the array. To wit, their approach to hardware maintenance and lifecycles, via the Evergreen Storage program, as well as their implementation of Pure1 has had me thinking for a while that they’re not your father’s AFA vendor.
I wrote about FlashBlade when it was first announced and I was cautiously optimistic that they were onto something kind of cool. As it happened, Pure spent a lot of time at SFD10 giving us a run-through on what some of the thinking around the design of the FlashBlade was, and it solidified some of my ideas around the product.
Some of the challenges Pure aimed to address when designing the FlashBlade was the need for scale in terms of:
- Capacity – from terabytes to petabytes;
- Concurrency – from a few users to thousands; and
- Access patterns – from small files and metadata to large, streaming workloads.
They also wanted to do this without drowning the users or administrators in complexity.
One of the key approaches to this problem was to adopt a modular architecture through the use of the blade chassis. While we talk a lot about the flash in Pure’s FlashBlade, the network architecture shouldn’t be underestimated. A key component of Pure’s “software-defined networking” is hardware (no, the irony is not lost on me), with two Broadcom Trident-II ethernet switch ASICs collapsing three networks (Front End, Back End and Control) into one high performance fabric providing 8 40Gbs QSFP connections into customer Top of Rack switches. This provides Pure with the use of a high performance, integrated fabric connected to scalable server nodes. While some of the specifications at the time of announcement were limited to the chassis, you’ll start to see these numbers increase as the SDN component is improved over time.
Brian was keen to see us thinking about the FlashBlade hardware design in the following terms:
- An integrated blade chassis provides density and simplicity;
- All-flash storage unlocks the parallelism inside an SSD; and
- An NVRAM engine built for distributed transaction processing.
Rob Lee then went on to talk about the software side of the equation, with the key takeaways from the software side of things being Pure’s desire to:
- Achieve scalability though parallelism at all layers;
- Create parallelism through deep partitioning and distribution; and
- Minimise the cost of distributed coordination.
Further Reading and Conclusion
Chris Evans did a nice article on Pure prior to SFD10. Chris Mellor did a decent write-up (something he’s prone to) at the time of release, and Enrico put together some interesting insights as well. Pure are certainly bucking the trend of commodity hardware by using their own stuff. They’re doing scale out differently as well, which is something some pundits aren’t entirely pleased about. All that said, I think the next 12 months will be critical to the success of the scale-out file and object play. Pure’s ability to execute on a technically compelling roadmap, as well as grabbing the interest of customers in rich media, analytics and technical computing will be the ultimate measure of what looks to be a well thought out product architecture. If nothing else, they’ve come up with a chassis that does this …