Speaker
Description
Regional traffic simulation supports mobility analysis and traffic forecasting by generating detailed vehicle movement data under different demand and road network scenarios. Producing these datasets requires large numbers of vehicles to be simulated and routes to be recalculated as traffic conditions change. Repeating the simulations with different settings creates two computational challenges: reducing the cost of individual route calculations, and processing large numbers of routing requests efficiently.
We use RUTH, a mesoscopic traffic simulator developed at IT4Innovations, to run regional simulations on high-performance computing (HPC) systems. Our Central Bohemian case features a road network of 194,585 nodes and 478,517 road segments and simulates up to 936,000 vehicles during the morning and afternoon period.
To reduce the cost of individual rerouting requests in this large use case, we are evaluating spatial and road class filters that restrict the road segments searched. Spatial filters limit the search to an area around the origin and destination, while road class filters exclude lower-priority roads from parts of the search. As an initial evaluation, we tested the filters on a representative sample of 50,000 origin–destination pairs from the Central Bohemian region. The ellipse filter reduced the median query time from 916.4 ms to 422.0 ms while still finding routes for 96.94% of the pairs. The hierarchical filter reduced the median query time further, to 107.1 ms, but found routes for only 86.38% of the pairs. This illustrates a trade-off. Tighter filters speed up queries but leave a small fraction unresolved, which still requires a less restricted search.
To process the large number of routing requests, RUTH uses the Asynchronous Communication and Execution library (ACE) to distribute them across HPC compute nodes through MPI or ZeroMQ. In weak scaling experiments on Karolina, we increased the workload from 1 million requests on 1 node to 16 million requests on 16 nodes while keeping the number of workers and requests per node constant. Elapsed time increased by only 1.8%, corresponding to approximately 98% weak scaling efficiency. Both backends therefore handled sixteen times as many routing requests with little increase in elapsed time when given proportionally more compute nodes.
These results address the two sides of the routing challenge. They reduce the cost of individual queries while scaling the overall workload with the available compute nodes. Next, we plan to implement and validate less restricted fallback searches for unresolved queries, evaluate the filters within full regional simulations, and assess how well the combined approach scales to larger geographical areas and vehicle populations.