Thermal-Aware Task Placement with Dynamic Thermal Model in an Established Datacenter

Abstract

Cooling system is one of the key components in data centers and consumes nearly half of the total energy. For the safety and efficiency of a data enter, cooling system should guarantee all the servers running in a suitable temperature. Thus, the cooling efficiency becomes lower as the occurrence of hot spots, because the whole cooling system will accordingly work in a strong cooling mode, which leads to overcooling. One of the key challenges toward this problem is to build an effective and efficient online task scheduling to balance the inlet temperature of the servers in a data enter. In this paper, we investigate the thermal model in a data enter, and we formulate the task placement problem to an optimization problem with the purpose to minimize the maximum inlet temperature of all servers. To get higher accuracy, we proposed a dynamic thermal model updated with the real data from temperature sensors. We solve the formulated problem in an approximate way and design a first-fit decreasing task placement algorithm with the idea of CPU budget. Finally, the effectiveness and accuracy of our algorithm are confirmed by experiments on a real test bed.