LB_Migrate: A Dynamic Load Balancing Library

The design of a general-purpose dynamic load balancing library for a vast variety of parallel applications is more challenging than the design of a static partitioning library. The dynamic load balancing library needs to be implemented in parallel with the application and must utilize memory efficiently, so that the application scalability is not affected. This paper studies the need for a dynamic load balancing library and looks at the related work. We propose a new dynamic load balancing library called LB_Migrate, targeted for large scientific applications with parallel loops as a major source of concurrency. The library is designed to be independent of any application. Hence, it must be supplied with a routine in the application that encapsulates the computations for a chunk of loop iterates. The library consists of various scheduling methods and load balancing strategies for performing both scheduling and data migration. Performance evaluation on a Linux cluster indicates that the library reduces the cost of executing an irregular loop without load balancing by up to 68%.