GPU Algorithm for Earliest Arrival Time Problem in Public Transport Networks

TL;DR

This paper introduces a GPU-based parallel algorithm for the earliest arrival time problem in public transport networks, employing novel pruning and data compression techniques to significantly outperform existing algorithms.

Contribution

It presents a topology-driven parallel algorithm utilizing pruning, clustering, and arithmetic progression techniques for efficient EAT computation on GPUs, achieving substantial speedups.

Findings

Achieves up to 59.09x speedup over CPU-based connection-scan serial algorithm.

Attains 12.48x speedup compared to existing GPU-based parallel algorithms.

Effectively processes large real-world public transport network data.

Abstract

Given a temporal graph G, a source vertex s, and a departure time at source vertex t_s, the earliest arrival time problem EAT is to start from s on or after t_s and reach all the vertices in G as early as possible. Ni et al. have proposed a parallel algorithm for EAT and obtained a speedup up to 9.5 times on real-world graphs with respect to the connection-scan serial algorithm by using multi-core processors. We propose a topology-driven parallel algorithm for EAT on public transport networks and implement using general-purpose programming on the graphics processing unit GPU. A temporal edge or connection in a temporal graph for a public transport network is associated with a departure time and a duration time, and many connections exist from u to v for an edge (u,v). We propose two pruning techniques connection-type and clustering, and use arithmetic progression technique appropriately to process many connections of an edge, without scanning all of them. In the connection-type technique, the connections of an edge with the same duration are grouped together. In the clustering technique, we follow 24-hour format and the connections of an edge are partitioned into 24 clusters so that the departure time of connections in the i^{th} cluster is at least i-hour and at most i+1-hour. The arithmetic progression technique helps to store a sequence of departure times of various connections in a compact way. We propose a hybrid approach to combine the three techniques connection-type, clustering and arithmetic progression in an appropriate way. Our techniques achieve an average speedup up to 59.09 times when compared to the existing connection-scan serial algorithm running on CPU. Also, the average speedup of our algorithm is 12.48 times against the parallel edge-scan-dependency graph algorithm running on GPU.