Exploring accelerator and parallel graph algorithmic choices for temporal graphs

Abstract

Many real-world systems utilize graphs that are time-varying in nature, where edges appear and disappear with respect to time. Moreover, the weights of different edges are also a function of time. Various conventional graph algorithms, such as single source shortest path (SSSP) have been developed for time-varying graphs. However, these algorithms are sequential in nature and their parallel counterparts are largely overlooked. On the other hand, parallel algorithms for static graphs are implemented as ordered and unordered variants. Unordered implementations do not enforce local or global order for processing tasks in parallel, but incur redundant task processing to converge their solutions. These implementations expose parallelism at the cost of high redundant work. Relax-ordered implementations maintain local order through per-core priority queues to reduce the amount of redundant work, while exposing parallelism. Finally, strict-ordered implementations achieve the work efficiency of sequential version by enforcing a global order at the expense of high thread synchronizations. These parallel implementations are adopted for temporal graphs to explore the choices that provide optimal performance on different parallel accelerators. This work shows that selecting the optimal parallel implementation extracts geometric performance gain of 46.38% on Intel Xeon-40 core and 20.30% on NVidia GTX-1080 GPU. It is also shown that optimal implementation choices for temporal graphs are not always the same as their respective static graphs.