Design and Evaluation of Dynamic Load Balancing Schemes under a Fine-grain Multithreaded Execution Model

The evolution of computer systems based on fine-grain multithreaded program execution models introduces both unique opportunities and tough challenges for the support of dynamic load balancing. Although load balancing is an active research topic in the distributed computing field, there is still a lack of a detailed study of the different dynamic load balancing strategies under a fine-grain multithreaded execution environment. This paper describes the design, implementation and performance evaluation of nine dynamic load balancing algorithms running on the EARTHSP multithreaded multiprocessor testbed a portable implementation of the EARTH multithreaded program execution model [4] on the IBM-SP2 multiprocessor system. In the course of this study we developed a set of generic test cases, which we call stress tests, that measure the performance of the different dynamic load balancing algorithms for specific workload patterns. Based on the experimental results from the stress tests and popular benchmarks, this paper shows that significant performance improvements can be achieved by using a dynamic load balancing algorithm that keeps track of adequate history information and accordingly employs both receiver-initiated and sender-initiated strategies to distribute the workload.