NPAC Technical Report SCCS-683

Multithreated Model for Load Balancing Parallel Adaptive Computations

Nikos Chrisochoides

Submitted November 20 1994


Abstract

We address the dynamic load-balancing problem of adaptive PDE computations on distributed-memory MIMD machines and on clusters of workstations (multicomputers). We present a multithreaded model for developing efficient and scalable algorithms and portable software needed to support at runtime the load balancing and communication requirements of adaptive methods on multicomputers. Our model is based on data-parallel and task-parallel paradigms. Task migration is based on a consumer-initiated Consumer/Producer paradigm which allows tasks required in load balancing to be overlapped with the tasks of performing the actual computation and communication required by the application. Under certain conditions, on the attributes of our model we can claim a better performance than that obtained by other existing load-balancing approaches.


PostScript version of the paper