Incorporating Latency in Heterogeneous Graph Partitioning

Parallel applications based on irregular meshes make use of mesh partitioners for efficient execution. Some mesh partitioners can map a mesh to a heterogeneous computational platform, where processor and network performance may vary. Such partitioners generally model the computational platform as a weighted graph, where the weight of a vertex gives relative processor performance, and the weight of a link indicates the relative transmission rate of the link between two processors. However, the performance of a network link is typically characterized by two parameters, bandwidth and latency, which cannot be captured in a single weight. We show that taking into account the network heterogeneity of a computational resource can significantly improve the quality of a domain decomposition obtained using graph partitioning. Furthermore, we show that taking into account bandwidth and latency of the network links is significantly better than just considering the former. This work is presented as an extension to the PaGridpartitioner, and includes a model for estimated execution time, which is used as a cost function by the partitioner but could also be used for performance prediction by application-oriented schedulers.