planning and decision making, and complex system design and analysis, including both hardware and Parallel Discrete Event Simulation (PDES) is an software systems. important application in use in many DoD projects; for In PDES, a simulation model is partitioned across example, PDES is used in large-scale war-gaming, and in several logical simulator processes (or LPs). Each LP complex system design, analysis and verification. processes its events in time-stamped order. Improving PDES performance

more »

... and capacity allows faster Synchronization among different LPs may be achieved simulation times and more extensive analysis of more using one of two major approaches: (1) Conservatively: detailed models. These benefits are not application-an event at an LP is processed only if all other LPs specific: they should reflect to any application that uses guarantee that it can be processed safely (no events earlier the improved simulation kernel. In this work; we than it will be generated to that LP); (2) Optimistically: overview our efforts for optimizing PDES in a LPs process events without concern for causality. Events Heterogeneous High Performance Computing (HHPC) received from other LPs, with a time stamp earlier than environment. We profile the SPEEDES simulator and the current simulation time, signal a causality error. Such identify several opportunities. We report on our errors are recovered from by rolling back the local experiences on two fronts: (1) optimizing the simulation state to a time earlier than the received communication subsystem--a critical system for PDES straggler event messages that were sent out erroneously. since it is afine-grained application and (2) exploring the To be able to achieve this synchronization, each LP must use of augmented FPGA boards to accelerate simulation. periodically checkpoint its state and event information. While such approaches have been attempted for Checkpoints are garbage collected when they are no sequential and data path intensive applications, we longer needed (when the global simulation time has believe that their use in clustered environments is novel passed them). This requires computing the Global Virtual Both efforts are works in progress; we present our Time (GVT) of the simulation to determine which history designs and some preliminary analysis results. For information may be garbage collected. Fujimoto l i l wrote example, removing the centralized communication server an excellent survey on PDES and PDES optimization from event message exchange path with a number of other approaches. small improvements to the simulation cycle, improved We describe our efforts in exploiting a performance by an average of 20% performance Heterogeneous High Performance Cluster (HHPC) to improvement for one of our large benchmarks. accelerate the performance of a PDES simulation engine. We start with the SPEEDES simulator: a state-of-the-art