Research On Key Issues On Energy Proportionality For Data Centers

With the development of information technology, more and more data need to be stored and processed, which leads to the rapid increasing of the scale and number of data centers. This in turn introduces more energy consumption in data centers. In general, the total energy consumption of a data center has made up a quarter of total operation cost. As data centers will not always operate in peak power with full workloads, reserarchers propose the "energy proportionality" concept, which means that the energy consumption is proportional to the workload changes. According to this idea, we can turn off some servers to save energy in case that the workloads towards the servers are in a low level. In order to satisfy the performance needs under heavy workloads, we should turn on additional servers. It is expected that we can greatly save the energy consumption of data centers if we can realize energy proportionality of data centers. Thus, it is valuable to study approaches to realizing the energy proportionality for data centers.Generally, data centers include server-based clusters which involve many servers and storage devices. So we mainly focus on the energy proportionality of server-based cluster systems and disk storage systems in data centers. The key issues that our paper focuss on are the energy proportionality of storage systems and server-based clusters. For the energy proportionality of storage systems, we focus on how to dynamically switch power states according to workload changes. Because data center usually show skewed access patterns, we divide the whole data into two parts:cold data and hot data, and the disks storing cold data are turned off. For the energy proportionality of server-based clusters, we focus on how to dynamically schedule tasks and switch power states based on workload changes.The main contribution of this paper can be summarized as follows:(1) We propose an energy saving method based on hot and cold grouping and dynamically reorganize file. This method divides total disks into a cold disk array and hot disk array based on history information. We can concentrate hot files into hot disk array by file migration and switching, so we can disclose cold disk array to save energy. In order to reduce the cost of file migration and switching, we design a method which can reorganize file distribution base on the change of access pattern. We introduce the ratio of access of hot file and disk to measure the diversification of file access pattern. We do the simulation experiment based on9synthetic trace and2real trace. The results show that our method outperforms current energy saving methods.(2) We propose an energy proportional method for server-based clusters, which is based on workload prediction. We get workload information in a time window, then we find the key points by a linear sequence fitting algorithm. We use the least squares linear algorithm to determine key points and predict the future workload, so that we can realize energy proportionality of server-based cluster by making smart decisions. We build a simulated server-based cluster prototype consisting of10nodes and conduct experiments to measure the performance of our proposal. Compared with the traditional method without energy proportional controls, our method performs better both in energy saving and in time performance.