Load Balancing Design Issues on Prefetch-Based DSM Systems

In recent years, the cluster computing technology has become a cost-effective computing infrastructure, because it aggregates resources of computational power, communication and storage. It is also considered a very attractive platform for low-cost supercomputing. Software distributed shared memory (DSM) provides a convenient and effective solution for programming parallel applications. However, both page faults and communication are major sources of overheads in DSM systems. Prefetching strategy can overlap data transporting time with computation time, as also reducing page faults. Unfortunately, it conducts load imbalance during barrier synchronization. For solving such inconveniences, this research paper discusses the load balancing for barrier synchronization in DSM systems. We discuss that, leaving the loop when half of hosts have finished prefetching is the best method, and therefore, we modify the threshold of leaving loop. Experiments show that, by incorporating load balancing into DSM systems, the barrier synchronization has been improved.