Performance Optimization for X10-based Large-scale Traffic Simulation

We are intensively working on the performance optimization for large-scale traffic simulation based on X10. What we are firstly facing is the synchronization overhead. If computational loads are not sufficient for given compute resources, the performance scalability can not be obtained due the the simulation model in that every synchronization is occurred at each step. The "step" here corresponds to "second". If the target road network is not that large, that synchronization can not be ignored. However, if we need to target larger network, such synchronization needs to be treated more carefully since all the cross points should be synchronized at every "steps". We introduce more relaxed simulation model in that the synchronization is occurred with longer interval. Determining the appropriate interval time is greatly important and the tradeoff between the interval and simulation precision exist. Thus we are currently working what's the most appropriate interval by given two facts, underlying execution environment and the precision threshold level. If we could find such a good balance, we will definitely be able to achieve more performance for given compute resources. The simulation itself had not been out of our research scope, but a lot of graph techniques could be leveraged for the performance optimization. Anyway, let's go on and publish this exciting research soon to international conferences.

X10 ベースの大規模交通シミュレーション実行基盤

道路ネットワークを領域分割することによって複数ノードで高速化されることは明らかであり、TSUBAME 上での妥当な結果が出つつある。ただ、一部の領域においては非常に時間がかかっており、これは車両の経路が通る重要な道路の許容量が少ないから。現在出ている性能分析により、今後は、より研究としての落としどころをつめていかないといけませんが、「Work Stealing 機構によって、動的に混雑度の高い領域に X10 アクティビティを生成する方式」を実装することにより、研究としての価値が生まれるでしょう。