| Summary: | 碩士 === 中華大學 === 資訊工程學系碩士班 === 92 === Cluster systems are becoming a mainstay of high performance computing. As traditional parallel processing, load imbalance exists in cluster architecture and it results in lower performance. There are some job scheduling systems presented with load balancing function, however, these algorithms were not designed for HPC PC cluster system absolutely, and there some defects can be improved. This thesis proposed a job scheduling scheme with two-phase load balancing architecture. It provides high performance load balancing capability and the two-phase load balancing scheme can effectively reduce extra overhead. We also implement a Linux PC cluster system with two-phase dynamic load balancing function and compare the performance with other load balancing algorithms. The experimental results show that our proposed algorithm has better performance on throughput and its average turnaround time is shorter than others. The throughput increases more than 20% and the average turnaround time decreases than 25% at least. Besides, we experiment with the impact of dynamic/static threshold and symmetric/asymmetric threshold schemes. It shows dynamic threshold can perform better than static one, and it can suit with light and heavy status. The asymmetric threshold scheme can make symmetric algorithm perform better than symmetric threshold scheme.
|
|---|
