Research And Implementation Of Dynamic Load-balancing Method Under Parallel Computing

Parallel computing is an effective way to achieve high performance computing. After the decades of development, the parallel computing has been widely applied in the field of engineering design, scientific computing, computing system, etc. Based on existing hardware, the problem of load balance is a research highlight to improve efficiency of parallel computing. And the dynamic load balancing method is in the advantages of flexibility and handling irregular problem, so that it is more and more attention.The dynamic available resource and delay both have influence on the efficiency of dynamic load balancing method in the parallel system. To the problem of dynamic available resource, the cluster job management system can manage parallel tasks better. Support for parallel computing has been integrated into majority of cluster job management system, but it is lack of consideration in the regard of task partition based on dynamic resource. What is more, to the delay influence, some foreign scholars research the delay load balancing further, and they propose corresponding validation methods. However, these methods all have some deficiency in the respect of time varying delay and system scale expandability. So, in this thesis, the dynamic load balancing method is researched and implemented from these two aspects.In this thesis, research focus on support for parallel computing of cluster job management system and delay dynamic load balancing. Firstly, by deep analysis of source code in Torque-the cluster job management system, the dynamic partition method is inserted in Torque, with the purpose of improving the efficiency of parallel tasks, system balance and overall utilization. Some corresponding experiments are done, and the result is verified that the proposed method is effective.Secondly, in this thesis, the delay dynamic load balancing model is proposed based on feedback control theory, and the computational method of optimal load balancing gain value is given. Then, the delay load balancing system is simulated by the discrete event simulation program. Comparing and analyzing the optimal load balancing gain value under the different conditions, the validity and effectiveness of the model has been verified. Using the different partition methods, the result indicate that delay affects system load balancing efficiency, by comparing respective system balancing time.