Parallel Discrete Event Simulation (PDES) is the parallel simulation of physical systems that may be modeled as a series of discrete events. Such events are fine-grained and often interdependent, which poses a challenge for efficient parallel simulators. An often-employed technique for uncovering parallelism in PDES is the optimistic execution of events without regard for causality constraints. Such protocols are called optimistic protocols. Perhaps the most well-known optimistic protocol, Time Warp, aggressively executes events, rolling back to a previous state and redoing computation when a violation of causality is detected. Because violations of causality represent the primary source of inefficiency in Time Warp simulations, substantial research in the field of PDES has focused on techniques for controlling optimism and reducing rollbacks. This work extends that research to achieve optimism control by means of Dynamic Voltage and Frequency Scaling (DVFS), a power management feature found in modern microprocessors.
This thesis examines the application of DVFS-based techniques to optimistic PDES simulations in order
to decrease simulation execution runtimes and power consumption. Adaptive protocols for optimistic PDES simulations are reviewed, as are existing DVFS-based techniques for power management and performance enhancement. A new adaptive optimism control protocol using DVFS algorithms in TimeWarp is proposed. This protocol is implemented in C++ and tested on two shared memory machines and a 20-node Beowulf cluster. Simulation performance is measured and energy consumption is estimated for each algorithm under several balanced and imbalanced workloads. It is shown that DVFS can be used to significantly reduce performance and power consumption in optimistic PDES simulations.