Thermal-aware Task Scheduling In A Datacenter

Some data shows that the expenditure of energy supply is becoming the major ex-pense of a datacenter. How to reduce unnecessary energy cost and improve the energy efficiency of computing equipment, reduce energy costs throughout the datacenter, re-duce greenhouse gas emissions has become a major problem and a main concern of the next cloud computing datacenter industry. Driven by all these, this article mainly talks about how to save the energy by task scheduling of the computing system in a data cen-ter. And we mainly solve the cooling energy wasting problem caused by imbalanced heat distribution.Firstly, we study the background and related work of the problem. First, through the description of heat recirculation and hotspot in datacenter, we get the reason for the inefficiency of the cooling system in current datacenter. Then we introduce the basic concepts and the related work of thermal-aware task scheduling. Finally, we focus on the two main steps while doing thermal-aware task scheduling. The first step is to build a power thermal model which is the basis of thermal-aware task scheduling. During online task scheduling, on one hand, we can update the dynamic power thermal model with the real-time inlet temperature from the temperature sensor, so we can make the prediction error of the model as small as possible; On the other hand, we can do task scheduling with this dynamic model to predict the temperature of the inlet temperature of all computing devices. The second step is to do thermal-aware task scheduling in datacenter. In this section, we focus on two aspects:task placement and task migration. While doing task placement, with the idea of CPU budget and first-fit descending (FFD) task placement method, we design an online thermal-aware task placement algorithm based on the dynamic power thermal model. We still use the dynamic power thermal model during task migration. We consider both the number of migration tasks and maximum inlet temperature while designing the online task migration algorithm.In order to verify the effectiveness of our model and methods, we design and complete a set of experiments on a simulated datacenter test bed. Experimental results show that our dynamic power thermal model has better prediction accuracy and the task placement and task migration approach based on this model can bring a smaller maximum inlet temperature, thereby saving more cooling system power. So the validity of our approach has been verified.Finally, we summarize our work in this paper and do some prospects on the future work.