A Worksation-Based Parallel Direct-Execution Simulator

It is important to understand and eeciently predict the behavior of large message-passing parallel programs executing on massively parallel machines, especially as these machines are becoming a dominant force in the scientiic computing arena. One important tool for this kind of research is direct-execution simulation, which provides performance and scalability analysis of application codes executing on top of some virtual computer of interest. We have developed such a tool called LAPSE (Large Application Parallel Simulation Environment), which is implemented on the Intel Paragon and provides excellent performance predictions for a wide variety of codes executing on a virtual Paragon. However, massively parallel machines such as the Paragon are scarce, expensive, and generally unavailable to large segments of the research community, making it important to implement such tools on top of computational platforms more widely available to the research community at large. To meet this need, we have ported LAPSE to a workstation cluster using the nx lib message passing library, which provides the function-ality of a Paragon on a network of workstations. The goal of this research is to provide researchers the opportunity to study the behavior of large parallel codes, written for massively parallel machines such as the Intel Paragon, while physically executing on a workstation cluster. This research is being carried out in two phases. The rst phase, which is now complete, involves improving the performance of LAPSE such that it is feasible to execute large simulations on a workstation cluster. In this paper, we describe our technique for reducing execution costs, and demonstrate up to a four fold increase in performance. The second phase, which is now in its infancy, involves investigating mechanisms to augment LAPSE on a workstation cluster with temporal capabilities: the ability to predict execution times for a code had it been executed on the Paragon. While we have no results to present as of yet, we do describe the approach we intend to pursue to solve this diicult problem.