Blocks and Files Facebook's Open Compute Project (OCP) could potentially destroy the business models of the mainstream server, storage and networking vendors. It is an attempt to turn back time and return to a pick-'n'-mix server, storage and networking world where suppliers happily build to common interfaces, and component sets aren't integrated for greater overall efficiency.
Facebook is its own massive systems integrator, but the rest of us are not and OCP standards adoption by enterprises could lead to much more expensive IT component acquisition and implementation processes.
How do I arrive at this point, a point where I'm trashing the spread of Facebook's wonderful OCP ideas?
OCP is Facebook's gigantic attempt to get the IT industry to deliver components in ways that suit Facebook; it wants the speed of hardware innovation to get closer to that of software by disintegrating rack and server components so that they can develop independently but still work together through common interfaces.
In this way, server CPUS and memory and interconnects and power supplies shouldn't have to develop at a four- or five-year refresh cycle rate with all the components in balance. Instead, for example, one could separate out the power supply boxes and have them supply all server components from their own shelf in the rack instead of being integrated with the server. One could have memory and processors loosely linked via Intel's new Silicon Photonics cable with 100Gbit/s bandwidth and very low latency - so you can't replace either CPUs or memory without disturbing the other.
Do this and Facebook can buy in its server components from cheap Taiwanese Original Design Manufacturers (ODMs) instead of complete servers from Cisco, Dell, HP or IBM. Facebook showed a Dragonstone database server at the Santa Clara OCP conference yesterday. It's for handling hot data and is built from - and uses - 3.2TB of Fusion-io PCIe ioScale flash card memory; there are no disk drives in it at all.
Elsewhere in the OCP data centre vision, Facebook is talking about using big and fat and slow disk drives with spin-down capability for storing cold data, like the photos that very few people will ever access again. It'll want these in the OCP rack design, not the industry-standard 19-inch rack design that enterprises use. And it won't buy OCP-standard cold data disk vaults from EMC or NetApp, preferring the ODMs again; they're cheaper and it won't have to pay EMC or NetApp data management software costs as it produces that software itself.
What Facebook wants is reasonable for itself but unreasonable for everyone else. If you are one of OCP's chosen vendors - oh lucky, lucky Intel if Silicon Photonics takes off - then riches and Facebook-blessed glory await you... as Fusion-io is hoping. But if you are out in the cold, with products like servers or storage arrays or network boxes, then OCP is a dagger aimed straight at your business model's heart, and blood is dripping from it already.
The now unspoken marketing aim for mainstream server, storage and networking vendors is to prevent their customers adopting OCP standards and ideas. Heaven forfend!
At the same time they have to play nice with Facebook and be OCP supporters, experimenting with producing their own stripped-down OCP-style product but only selling it to Facebook-style customers while everyone else prefers the full-fat products with full-fat profit margins.
Note that storage suppliers EMC, Fusion-io, Hitachi GST and Sandisk have all joined the OCP organisation. Fusion sees big bucks coming from its membership - ditto SanDisk I dare say, also also Hitachi GST with helium-filled big fat disk drives being hopefully adopted by OCP. But EMC? Why the hell has it joined? Know your enemy, I think.
If Facebook was buying millions of cars a year then yes, we could see it would want to separate engine development from that of the clutch, the transmission, the suspension, the braking and so on to speed its ability to use the latest and greatest components. The rest of us are different. We buy cars in ones and twos, tens and 20s if we have a fleet, and don't want to integrate the bits ourselves. We want a buy and drive a vehicle, not a buy and build box of bits.
That's why the OCP IT model will fail in the real world outside the hyperscale data centres of Amazon, Facebook, Google, Yahoo and the other enormous cloud IT service operators.
Don't believe me? Shout out why in the Reg OCP forum topic. ®