Optimum control method of workload placement and air conditioners in a GPU server environment

High-performance computing datacenters have been rapidly growing, both in number and size. In addition to the conventional CPU servers, more GPU servers are being placed in datacenters to achieve high-speed processing such as image recognition. From our experiments, the power consumption during GPU operation is greatly affected by changes in the temperature of the server intake port. Therefore, to reduce the total power consumption of datacenters, thermal management of datacenters can address dominant problems associated with cooling such as the recirculation of hot air from the server outlets to their inlets and the appearance of hot spots. In this paper, we propose a workload placement method for environments where CPU servers and GPU servers coexist, and an optimum air-conditioning control method that cooperates with the workload placement method to reduce the total power consumption of the servers and air conditioners in datacenters. Experiment results in an actual machine environment showed that our proposed method has valid power-saving effects by adjusting cooling capacity tradeoffs between GPU servers and air conditioners.