Big data centre normally will use a lot of energy since the machines housed in it is running 24 hours and there are a lot of machines in the place. For the longest time, Google has designed and built data centres which use half the energy of an average data centre. Recently, they developed a new tool called machine learning to increase the efficiency of a data centre. They also have released the white paper (PDF) on how they are using neural networks to optimize data centres operations and decrease the energy usage of the data centre.
Google’s data centre’s engineer, Jim Gao, knows well on the daily operation of the data centre. From the daily operation data, Google do PUE calculation. It is a calculation to measure energy efficiency. The result is that for every 30 seconds, the data centre was tracking on total IT load (the amount of energy used by the servers and networking equipment at any time), outside air temperature (which affects Google’s cooling tower) and the levels which Google’s mechanical and cooling equipment was set.
Therefore with these data, Jim developed a model to analyze, recognize patterns and predict the next pattern and eventually it will help to improve the data centre performance. After a few testing rounds, Jim’s models are now reaches 99.6 percent accuracy in predicting PUE. For example, if a few servers were put down for a few days, it will use less energy usage which in the end will have impact for the PUE calculation and in the end Jim’s model can be use to change the temperature setup. These kind of small tweaks can help in saving both energy and money.