The Data Center Disaggregated: Promises and Problems

The software conundrum of best-of-breed components versus tightly integrated stacks is nearly as old as IT itself. In the last several years, the prevailing winds have blown heavily toward best-of-breed as the combination of software-as-a-service and increasingly flexible APIs has ushered in an era of new interoperability in the enterprise.

Now, thanks to Facebook's Open Compute Project, which seeks to make computing hardware more efficient and scalable, we're inching closer toward a similar reality for the data center. At OCP's most recent summit, held in January, the 50-plus members who are now part of the community proposed the latest step toward a truly modular data center: disaggregation.

More specifically, OCP (no relation to the ominously corrupt corporation that runs Detroit in1987's RoboCop) wants to see the core components of system design--processor, motherboard, network interconnects and so on--separated so that they don't depend on each other too heavily and can be upgraded independently.

"They've already done an open rack where everyone's equipment can fit in," says John Abbott, a distinguished analyst at 451 Research who authored a recently published research note based on what he heard at the January summit. "Now they want open motherboards."

Abbott believes OCP's idea could be a boon to data center operators, who are so desperate to minimize capital expenditures that they might be attracted to being able to separate motherboard components so that they can take advantage of the latest and greatest technologies on the market.

"To upgrade CPUs to the latest high-performance model would in the past have required a whole new pizza box," Abbott wrote in his research note. "With disaggregation, the upgrade process potentially becomes a lot more flexible."

Before the dream of disaggregated systems can be realized, a number of obstacles must be negotiated. For one, there are missing pieces to the puzzle, such as a fast interconnect between components, and more widely adopted standards. The former may be addressed by ongoing efforts--led by OCP member Intel--to develop optical interconnects called silicon photonics, which would enable light to carry data at very high speeds while consuming very little power. Abbott says the first such products are expected later this year.

OCP's vision also must overcome system vendors' hesitation to tear down the proprietary walls that have fed their profits for so long. It's a daunting task given the relatively limited market potential--at least at first--for modular system components.

"Buying systems at a more granular component layer isn't for everybody," Abbott wrote in the research note. "Only the largest data centers with their own systems integration capabilities will be able to take full advantage, at least until systems vendors adapt their business models--which they may be reluctant to do."

Which leaves us with a classic Catch 22: For large-scale adoption of modular systems to occur, vendors must agree to design components that interoperate with their rivals' products; meanwhile, vendors are likely to want evidence of a large-scale market opportunity before they'll make the required investments and loosen up their proprietary ways.

While it may take time to resolve that conflict, it's clear that OCP's voice is getting louder, and that vendors' ears have at least perked up.

"OCP is gathering real momentum, with user adoption taking off and vendors keen to participate because they see some real business opportunities emerging," Abbott wrote in his research note. "The open specification model and projects voted on by users gives OCP the chance to explore standardization areas where vendors by themselves will not go."