According to a recent article in GigaOM, server underutilization is the dirty little secret of the data center world. In 2008, a report from McKinsey showed that data center utilization was hovering around 6 percent and, last year, Gartner put utilization rates at 12 percent.
The utilization problem was supposed to have been solved by virtualization technology. By splitting physical servers into multiple virtual servers, the hardware can be used at peak capacity, or so the argument goes; but, in reality, we’re still not making the grade when in comes to fully utilizing server resources, particularly CPU power.
Internet giants like Facebook are building mega-data centers everywhere from Altoona to the Arctic Circle, but more of the same doesn’t solve the basic problem of underutilization, which is caused in large part by the design of servers. A bigger-is-always-better mentality has shaped a chip manufacturing industry that is always in pursuit of more processing power per square inch of silicon. That sounds good, but in reality, current processing speeds outstrip potential data I/O by a significant margin. Processors are sitting idle because servers can’t push data fast enough to feed them.
One potential solution can be found in ARM SoCs. These systems on a chip are generally much less powerful than their Intel siblings, but that’s no bad thing. If we want to create servers that can be maximally utilized, the smart thing to do is to match up processing speeds with anticipated workloads and with data throughput. Using clusters of lower powered (and less power hungry) ARM SoCs makes it easier to build servers suited to their particular tasks.
In the same article, figures supplied by Google reveal that their three-month average load on a 20,000 node cluster was between 20 – 40 percent. That’s perhaps part of the motivation for the recent news that Google is considering making the switch to ARM-based servers, news that saw Intel’s share price drop by 3% — Google makes up more than 4% of the chip giant’s revenue. By bringing chip design and manufacture in-house, Google will no longer have to rely on one-size-fits-all overpowered chips and can create a hardware infrastructure that meshes with its workloads and software, allowing the company to build a platform that can be utilized at higher levels of efficiency.
Of course, most companies aren’t Google and don’t have the wherewithal to manufacture their own chips, but ARM SoC servers that can be used by web hosting companies and data centers are already on the market and — as we discussed in our recent article on Data Center Knowledge — in 2014 64-bit ARM SoC-based servers will hit the market, at which point we can expect to see data center owners paying serious attention to the way that ARM servers can ameliorate the underutilization problem.