Research Of Energy-Efficient Scheduling Algorithm Based Task Dependency On Homogeneous Clusters

The rapid growth of scientific data and commercial data has made a very big challenge to computer technology, and large-scale applications urgent need high-performance computing systems, mass storage systems and high-speed network infrastructure to support. In high performance computing area, researchers and designers have been striving for higher system performance, they tirelessly to increase the number of system processors and improve processor performance in pursuit of higher FLOPS(Floating-Point Operations Per Second). However, power management on processor cluster in a very long time has been neglected by them, until in recent years, has gradually been paying attention.In large computer systems, energy consumption on computing and cooling has occupied most of the total energy consumption. Substantial growth in energy consumption produced a lot of heat, which not only increase the cost and the difficulty of cooling, but also affect the stability of the system components. Meanwhile, computer-generated greenhouse gases on the environment adversely affect the formation and the gas showing an increasing trend. Therefore, the study on energy-aware schedule has a far-reaching significance.In the research of reducing energy consumption on cluster system, energy-aware task scheduling is a kind of ideal solution. Overall, the parallel task scheduling strategy can be divided into three types:priority based scheduling, cluster based scheduling, task-duplication based scheduling and random search based scheduling. Among them, the task duplication based scheduling has a better performance than the other two methods. However, the task duplication based schedule replicate tasks on the critical path to shorten the total time scheduling. If the redundant tasks required for the execution of energy than the communication energy consumption reduced, it will cause the system total energy consumption increase. Therefore, in order to ensure the scheduling system performance, energy efficient scheduling algorithm based on task duplication has practical significance.On the basis of summarizing previous work, a careful study of the existing task duplication based scheduling algorithm and energy optimization algorithm, I analyses the advantages and defects of existing scheduling strategy, proposed a task scheduling based on dependency degree. In this paper, I first proposed the inter-task dependency degree, task-path dependency degree and the path-processor dependency degree, and then get the task duplication scheduling algorithm by limiting the number of processors and selecting the most energy efficient processor for allocating tasks execution path, which can reduce the more energy consumption of the system, release more idle processors, and the load of processors more balance.In the end of this paper, I designed a platform for experiment simulation using C++, and implemented three improved algorithms and three task duplication based original algorithm. The experimental platform configure the parameters same to the processor and high-speed Internet network used in the actual, and execute the two common tasks set in cluster system for testing. In the experiment, I set the value of slack factor, the number of processors, the type of task sets, the type of high speed connect devices and the value of communication-computation rate in different tests, and get the data of schedule length, total energy consumption and the used processor count. Through the experiment, the improved algorithm can significantly reduce the used processor number, improve the processor load balancing and save more energy consumption; especially on communication intensive tasks and high network communication delay have a better effect.