The Research Of Energy Conservation Mechanism For Server Clusters In Network

With the development of the science and technology, computers are widely applied in many fields, and their performance is expected higher and higher. A possible solution is to introduce giant computers, to meet the large-scale scientific computing of there emerging industries. But the astronomical costs of the giant computers are high enough to discourage most of the potential users. As the large-scale server clusters bring great convenience for scientific research, work, study and daily life, it also brings another serious problem—a huge energy consumption issue. A power issue is an important problem for server clusters, because it is not only a concern of a single computer or a group of servers, but also directly influences their cooling requirements. In fact, a medium to large number of high-performance nodes racked closely in the same work environment, as is usually the case with clusters, requires a significant investment in cooling, both in terms of sophisticated racks and heavy duty air consumption also influences the required investments in backup cooling and backup power-generation equipment for clusters that can never be unavailable, such as those of companies that provide services on the Internet. By extension, the server cluster energy conservation becomes a major issue for the whole society. Although, many local authorities invested heavily in power plants to solve power problems, most power-generation technologies have a negative impact on the environment. Clusters energy-saving issue has been a very real, serious problem, not only from the perspective of the Internet, but also from the point of the whole society. There are some mature energy-saving strategies for system-level scheduling, such as dynamic power management policy (DPM), dynamic voltage and frequency adjustment strategy (DVFS), dynamic voltage scaling strategy (DVS) and so on.These strategies are applied successfully on system-level energy conservation, but they are not suitable for cluster level. The strategies for cluster energy conservation have become researching hot issues. PID and load concentration (LC) have been occurred for a long time. With the complexity of cluster structure and diversification of cluster business, these methods'effectiveness reduces.Dynamic cluster configuration is dynamically adjusting the size of servers based on the network load in order to achieve optimal service performance under the minimum system power consumption. This paper proposes a prediction-based dynamic clusters configuration strategy, it uses Least Mean Square (LMS) and Recursive Least Squares (RLS) to predict the situation of service requests in the future time according to the network historical information of service requests, then on the basis of the load requests and the clusters processing power to decide the servers scale and dynamically adjust the opening and shutdown of the computers in the server cluster.Besides, we proposed a particular cluster reconfiguration strategy for computation intensive service. In order to provide a quality-of-service in terms of overload probability, we formulate the problem of energy consumption as a constrained optimization problem, i.e. minimizing the number of active servers to reduce the energy consumption while keeping the overload probability below a desired threshold. An overload probability estimation model is derived by applying large deviation principle. Starting from this model, an online measurement based algorithm for dynamic cluster reconfiguration is developed to decide the number of servers to power on/off. The proposed algorithm makes decision based on current workloads without the requirement of prior knowledge of the workload statistics. Another advantage of the proposed algorithm lies in its implementation based on a type of iterative policy to adjust the active servers, instead of directly determining the number of active servers.In the simulation, we used the Parallel Workloads'trace data of Hebrew University of Jerusalem. And we use real user accessing data in cluster to simulate. And the simulation verifies the feasibility and advantage of the scheduling strategy.