Effectiveness of a Dynamic Load Balancing Library for Scientific Applications

The design of a general-purpose dynamic load balancing library for a vast variety of parallel applications is a very challenging task. The library has to address potentially unpredictable load imbalance in the application, to interface with the data structures native to the application and to achieve significant performance improvement and scalability. In this paper we look into the design and sample results of a new dynamic load balancing library called LB Migrate, targeted for large scientific applications with parallel loops as a major source of concurrency. The applicatons must supply the library a routine that encapsulates the computations for a chunk of loop iterates, and the data for the computations must be stored in an array of arbitary type. We demonstrate the effectiveness of the library on two real applications - the profiling of an automatic quadrature routine problem and for a simulation of a hybrid model for image denoising. The experimental results indicate that the library achieves up to 60 % performance improvement for these applications.