Evaluation of Mapping and Dynamic Load-Balancing Strategies in SEPP-HPCTI

A fundamental issue affecting the performance of a parallel application running on message-passing parallel systems is the assignment of tasks to processors in order to get the minimum completion time. In SEPP and HPCTI, tools for static and dynamic task assignment are considered complementary: static mapping tools compute and initial assignment of tasks on processors while dynamic load balancing tools are used at execution time. Static task assignment is known to be NP-complete and, therefore, the obvious approach is to develop polynomial time algorithms that provide fast and near optimal solutions. In this paper, we present a compilation-time two stage mapping strategy (denoted as CREMA: Clustering and Reassignment-based Mapping Algorithm) used for mapping efficiently arbitrary programs (modelled as TIGs: Task Interaction Graph) onto message-passing parallel systems with any architecture. Moreover, we present also a new fully distributed dynamic load balancing algorithm (denoted as DASUD: Diffusion Algorithm Searching Unbalanced Domains) for load balancing among the processors of an arbitrary interconnected network of processors. We present a description of both strategies and the results obtained in their respective evaluation. From these results, we conclude that CREMA and DASUD achieve better results than other static and dynamic load-balancing strategies proposed in the literature.