Scalable Optimistic Parallel Simulation

Computer simulation of large and complex systems in various elds of science and engineering requires tremendous computational resources, and is limited by sequential simulation algorithms in most practical cases. Parallel simulation ooers the potential to accelerate simulations so as to facilitate the study of complex dynamic systems with large problem sizes, to explore large parameter spaces, or to iteratively improve a parameter estimate. In this paper we give an overview of two main approaches in synchronizing parallel event execution, namely conservative and optimistic. The conservative approach executes safe events only, while the optimistic approach executes events without the consideration for event safety and uses a rollback mechanism to correct event causality errors. Speculative event execution adopted in the optimistic approach can potentially uncover a higher degree of parallelism in the simulated system. However, runaway speculation may result in excessive rollback recovery, and limits the scalability of the optimistic approach. We propose a two-sided throttling scheme that dynamically controls and adjusts the event execution speed in each simulation process. The throttle is based on a new concept called global progress window, which allows the individual simulation process to be positioned on a global time scale, thereby to accelerate or suspend their event execution. No global synchronization is required in the throttling scheme. Experimental results show that the throttling scheme reduces the number of rollbacks by 30%, and the number of executed events by 60%. The improvement is consistent for simulation with large problem sizes. Lastly, the scalability of the throttle is investigated.