Abstract
Many fields of modern science rely more and more on the immense computing power of supercomputers. Modern, multi-thousand node systems can consume megawatts of electrical energy in highly uneven manner, challenging the data center infrastructure, both power and cooling coils. The traditional way of managing the infrastructure makes each subsystem of a data center (e.g. cooling) independent from all other in the way it relies only on local sensors to manage the infrastructure. The erratic nature of computing in a large data center makes this approach suboptimal. In the paper we show that by challenging the traditional split between the infrastructure and the computing equipment, one can gain significant boost in energy efficiency of the entire ecosystem. A solution that predicts cooling power demand basing on the information from a supercomputer resource manager, and then sets up the parameters of the cooling loop, is presented along with potential benefits in terms of reduction of the power draw.