Facebook stretches ARM chips in datacentre tests

Facebook is putting ARM and Tilera processors through tests in its datacentres, as it explores using non-x86 chips for some applications to reduce its electricity bill and boost performance. 

Two hardware engineers at the social-networking company told ZDNet recently that they are testing the processors and are keen on them for certain applications. However, they did not disclose how many chips the company is using.

On a recent visit to Facebook's hardware lab, the company's manager of system engineering Amir Michael explained that the chip testing was going on in Facebook's datacentres.

"The reason you don't see any [equipment] here is that all of those ARM and Tilera systems are being tested against production traffic," Michael said. 

Facebook is evaluating ARM's chips for their low-power characteristics and reasonable compute capability. However, it won't be until the launch of 64-bit variants next year that it gets serious about the processors, the engineers indicated. 

"I think ARM-64 is much more interesting for datacentre computing," Matt Corddry, Facebook's manager of hardware design, said. "We're very interested about that hitting the market... we love what the ARM community is doing."

ARM for low power...

ARM is best known for making the processors that sit inside most of the world's smartphones, including Apple's iPhone 5. Its chips are used so widely because their RISC instruction set is somewhat simpler than Intel's x86, which helps them consume much less power than Intel's product. A Calxeda ARM-based server draws around 5W per server node, compared with around 20W for some of Intel's most energy-thrifty Xeon E3 server processors. 

When I asked Calxeda whether Facebook was using its ARM servers, the usually chatty start-up said, "No comment."

"I think ARM-64 is much more interesting for datacentre computing... We love what the ARM community is doing" — Matt Corddry, Facebook

Over the past few years, ARM believes the nature of servers has changed enough for its chips to have a chance at getting into big datacentres.

The shift has been driven by a growing appetite by consumers and businesses for cloud services, which has led to a boom in the construction of massive datacentres. This growth has forced cloud companies to pay ever-closer interest to their electricity bills, prompting interest in low-power ARM servers from the likes of Marvell and Calxeda.

For a cloud operator like Facebook, reducing power consumption can have a major effect on either how many servers they can field, or how large their electricity bill is. For example, the company's flagship datacentre in Prineville, Oregon, is expected to consume around 80MW once it is fully built. Assuming a power usage effectiveness (PUE) rating of 1.25, then this means it will have 64MW to spend on servers, network and compute resources.

At this scale, the difference between 5W per server node and around 20W has huge repercussions, as Facebook can fit four ARM servers into the same power profile as one Intel server. 

...Tilera for high performance

As for Tilera, the motivation here is less about power and more about efficiency. The company's 64-core processor demonstrated a 67 throughput gain over Intel's chips at handling a Memcached workload, according to a Facebook case study (PDF). 

"Facebook has openly talked about the advantages they see with the Tilera TILE-Gx processor and our lead over other non-x86 processors in having low-power, high-performance, 64-bit cores in production today," Tilera told ZDNet.

ARM in the cloud, Intel in the enterprise 

Intel and AMD's x86 processors are excellent chips for general purpose computation. However, as cloud companies have grown, they have found that they can eke out efficiencies from using different processors for known workloads.

The difference between a typical enterprise and a cloud company is that a business will run a different application on each server, or thereabouts, while a cloud company will "run the same application thousands and thousands of times across different servers", Michael said.

For this reason, cloud companies can reap major benefits by tuning certain applications to take advantage of different processor architectures. This can make the cost of porting and tuning application code less of a deal breaker.

"There's an opportunity to tune it, and then you're able to get very good [performance] on one application," Michael said.

For the time being, it seems that ARM and Tilera processors will be a useful tool in a large datacentre operator's technical arsenal, but until it gets easier to port code over, use will be restricted to companies with massive applications that span many servers.

Nonetheless, Facebook is sending a clear signal to AMD and Intel that they're not the only chipmakers in town anymore. When asked if he thought Facebook could put ARM workloads into production, Michael said: "It's not a question of if; it's a question of when."