Coho CTO on Pure's FlashBlade: Oh-ho, that's a no-go no-no
Design is wrong way to go, apparently
Comment Andy Warfield, Coho Data's cofounder and CTO, thinks Pure Storage's FlashBlade design is misconceived.
Warfield says Coho's engineering team went down a very similar route, found it to be wrong, and refocussed elsewhere.
What happened was that, in 2013, the team thought it a good idea to build the densest network-integrated hardware platform that it could to run the software Coho was developing. They were building around PCIe flash and saw that it needed more compute and connectivity than SSDs, and the compute should be separate from the flash because falling flash prices meant you should buy it separately from your compute hardware.
In fact compute, network and storage all needed to scale well, and Coho looked to pack all three into a single hardware chassis.
Their idea was to use an Intel Seacliff Trail 10Gbit, top-of-rack switch as a blade server backplane. Half the ports would interconnect storage blades in the proposed SwitchStore system, with the other half being used for external access devices and exposed as an Ethernet switch.
The storage blades (FlashBlade precursors of this had come to fruition) would be PCIe flash cards hooked up to a server card, and then to the Seacliff Trail backplane.
But the Coho engineers decided this elegant idea was in fact a design cul de sac, because for one thing, you simply should not bet against commodity hardware as, compared to proprietary designs, it has shorter cycle times, lower costs, and smaller challenges in getting validation and QA right.
Commodity hardware benefits enormously from the sheer volume of its deployment. Warfield says: "The idea of being stuck with the delivery lifecycle of proprietary hardware scared the crap out of us."
With software-driven products, you can focus on innovation there and take advantage of fast-developing compute, storage and networking commodity hardware as they occur.
A second point counting against the SwitchStore concept was that it can feed to so many servers that network access to it could be a choke point, involving much cross-rack storage traffic.
It's better, Warfield argues, to provide top-of-rack storage and keep the bulk of storage networking traffic inside a rack. He sees FlashBlade as being so powerful in storage terms that it could seriously affect a data center's core network performance with the amount of inter-rack traffic it generates.
He thinks the scale-out storage concept should include the flexibility to adapt to emerging hardware, to buy hardware as it's needed, and to maximize hardware's efficiency in your data center. These things run counter to the complexity of traditional enterprise storage systems, and Pure's FlashBlade is not really a scale-out array in his view.
It's "an expensive, proprietary, blade chassis-based array," while also being "a really interesting, impressively complicated, proprietary piece of enterprise storage hardware."
You can read his argument in more detail on Andy Warfield's blog. ®