Distributed, low contention task allocation

Designing a good task allocation algorithm faces the challenge of allowing high levels of throughput, so that tasks are executed fast and processor parallelism is exploited, while still guaranteeing a low level of memory contention, so that performance does not su er because of limitations on processor-to-memory bandwidth. In this work, we present a comparative study of throughput and contention guarantees provided by load balancing networks, a new class of distributed, asynchronous algorithms for real-time task allocation in shared memory multiprocessors. Load balancing networks generalize balancing networks, to accomodate tasks with varying completion times. On the theoretical side, we formulate precise and crisp de nitions for capturing the quality of load balancing provided by general task allocation algorithms; we use these de nitions for formally evaluating the throughput performance of speci c constructions of load balancing networks that we propose. Furthermore, we introduce a formal, complexity-theoretic measure of contention incurred by tasks with varying completion times, and use it to analyze the contention performance of these constructions. Our theoretical results display precise and subtle trade-o s between throughput and contention performances for load balancing networks. On the practical side, we propose an experimenThis work has been supported by ESPRIT III Basic Research Project # 8144 (LYDIA { Load Balancing on HighPerformance Parallel and Distributed Systems), and by funds for the promotion of research at University of Cyprus (research projects \Load Balancing Problems on Shared Memory Multiprocessor Architectures" and \Distributed, Parallel, and Concurrent Computations"). Department of Computer Science, University of Crete, Heraklion 71110, Greece, & Institute of Computer Science, Foundation for Research and Technology { Hellas, Heraklion 71110, Greece. E-mail: [email protected] Department of Computer Science, University of Cyprus, Nicosia 1678, Cyprus. Part of the work of this author was performed while visiting Institute of Computer Science, Foundation for Research and Technology { Hellas. E-mail: [email protected] tal platform for evaluating the actual performance of load balancing networks through a series of carefully designed experiments that simulate these networks on real shared memory multiprocessor machines. Our experimental approach encompasses a rigorous methodology for randomly generating tasks that are not merely \random", but rather belong to common classes of tasks such as periodic and sporadic. Our experimental results reveal that load balancing networks substantially outperform in performance classical, centralized methods for task allocation.