Quantitative Driven Optimization of a Time Warp Kernel

The set of events available for execution in a Parallel Discrete Event Simulation (PDES) are known as the pending event set. In a Time Warp synchronized simulation engine, these pending events are scheduled for execution in an aggressive manner that does not strictly enforce the causal relations between events. One of the key principles of Time Warp is that this relaxed causality will result in the processing of events in a manner that implicitly satisfies their causal order without paying the overhead costs of a strict enforcement of their causal order. On a shared memory platform the event scheduler generally attempts to schedule all available events in their Least TimeStamp First (LTSF) order to facilitate event processing in their causal order. By following an LTSF scheduling policy, a Time Warp scheduler can generally process events so that: (i) the critical path of the event timestamps is scheduled as early as possible, and (ii) causal violations occur infrequently. While this works effectively to minimize rollback (triggered by causal violations), as the number of parallel threads increases, the contention to the shared data structures holding the pending events can have significant negative impacts on overall event processing throughput. This work examines the application of profile data taken from Discrete-Event Simulation (DES) models to drive the simulation kernel optimization process. In particular, we take profile data about events in the schedule pool from three DES models to derive alternate scheduling possibilities in a Time Warp simulation kernel. Profile data from the studied DES models suggests that in many cases each Logical Process (LP) in a simulation will have multiple events that can be dequeued and executed as a set. In this work, we review the profile data and implement group event scheduling strategies based on this profile data. Experimental results show that event group scheduling can help alleviate contention and improve performance. However, the size of the event groups matters, small groupings can improve performance, larger groupings can trigger more frequent causal violations and actually slow the parallel simulation.