| Abstract |
Urban traffic simulations are of great importance for smart cities research. However, city-scale simulators can be both process and memory-intensive, and hard to scale. To speed up these simulations and to allow the execution of larger scenarios, this work presents a set of optimizations based on two complementary approaches. The first is an approach inspired by SimPoint to estimate the results of new simulations using previous simulations. This technique consists of identifying and clustering recurring patterns during a simulation, and then using a representative (the centroid of the cluster) of the recurring patterns to reconstruct the remaining ones. On a dataset with 216 time series, our technique was able to estimate the original series (of the same simulation) with an average error of 6.38 x 10(-6). Using only the trips which include the centroids (50% of the total simulation), estimation of metrics such as average speed and percentage of street occupancy rate, presented errors of 1.2% and 30% respectively, with a speedup of 1.95 in execution time. The second approach works on a lower level. In this approach we explore alternative implementations to Erlang's ETS tables, a central data structure used by the InterSCity simulator, and a current performance bottleneck. These optimizations yielded a speedup of approximately 1.28 when compared to the current version of the simulator. |
