Workshops Archive

Distributed SpMV with HClib-Actor: A Case Study in Alternatives to MPI-Based PETSc


Workshop: The 9th Annual Parallel Applications Workshop, Alternatives to MPI+X (PAW-ATM)

Authors: David Pang, Ajay Sharma Mandadi, Shubhendra Pal Singhal, Pulkit Gupta, Vivek Sarkar, Akihiro Hayashi, and Thomas M. Conte (Georgia Institute of Technology)

Abstract: Sparse matrix-vector multiplication (SpMV) exhibits irregular computation and communication patterns that challenge distributed-memory execution. This paper presents a case study of distributed SpMV using HClib-Actor, an actor-based programming and runtime system implementing the Fine-grained Asynchronous Bulk-Synchronous Parallel (FA-BSP) model. We explore four HClib-Actor-based implementations as alternatives to MPI, varying matrix partitioning and sparse storage format, which together determine the communication direction and local traversal order. We evaluate their weak and strong scalability on synthetic and real-world graph matrices using up to 1,536 CPU cores, with PETSc as a highly optimized MPI baseline, on the Georgia Tech PACE Phoenix cluster. Our results show that HClib-Actor achieves good scalability on relatively well-balanced workloads and outperforms PETSc in several configurations, although performance remains sensitive to workload distribution. Overall, this study demonstrates the potential of HClib-Actor for distributed SpMV and motivates further exploration of partitioning, scheduling, storage, and communication strategies.


Back to The 9th Annual Parallel Applications Workshop, Alternatives to MPI+X (PAW-ATM) Archive Listing

Back to Full Workshop Archive Listing