Earlier this week at Data Centers Europe 2014, Joe Kava, Google’s Vice President of Data Centers, announced that they had begun employing a “neural network” built by another Google employee, Jim Gao, that has the ability to use monitoring data to create a 3D model of airflow in a server room. Gao’s goal was to track a large set of variables in order to ultimately determine efficiency – variables like distributed server load, weather conditions, cooling tower operations, water pump and heat exchanger operations, etc. Google says that this “neural network”, as they’re calling it, is now accurate to 99.6% at predicting Google’s Power Usage Effectiveness (PUE).
At a time when more data centers than ever before are calling into question the necessity of mechanical cooling and whether there exist other, better methods of cooling. Some data centers, like QuadraNet, are turning to air-side economizers, which synchronize outside temperatures with inside data center temperatures. These tools aren’t permanent solutions, since ASHRAE recommended operating ranges are between 18C and 27C (65F and 81F), which means that on a particularly hot day, it may be 90F+ outside, and it would obviously not be beneficial to turn an air-side economizer on (in fact, quite the opposite, you’d probably end up with a number of failed machines). While newer hardware is built to higher standards and is better able to withstand higher temperatures, older hardware can be extremely prone to failure at temperatures above ASHRAE’s recommended operating range. Since a large percentage of data centers built for public use, (unlike Google, which owns and operates all of its own servers), still use some legacy hardware, it seems that mechanical cooling is here to stay, at least for the foreseeable future. Perhaps the introduction of newer equipment that supports IPv6 will aid in allowing data centers to rid themselves of some of this older, legacy hardware, which will in turn allow them to eventually move away from mechanical cooling.
It’s unclear whether Google intends to release Jim Gao’s software or “neural network” for public use, though it seems likely they may keep it internal, at least for the time being while they can still reap its benefits. One thing is for sure though – if data centers around the world were able to make use of this technology, it’s a solid bet that global commercial power consumption would fall and data centers would run substantially more efficiently.