A Benchmark of C++ Cluster-Computing Libraries: MASS C++, HPX, and PM2

Abstract

The selection of a distributed C++ runtime for high-performance cluster applications entailssignificant performance and programmability trade-offs, yet existing evaluations of higher-level runtimes remain siloed within each runtime’s own application domain, precluding workload- diverse, controlled cross-runtime comparison. This thesis presents a benchmark suite of five computationally distinct patterns—a 3D stencil kernel (Heat3D), dense matrix multiplica- tion (DGEMM), graph motif search, a phased financial graph simulation (Bail-In/Bail-Out, hereafter BIBO), and an agent-based model (SugarScape)—implemented equivalently across MASS C++, HPX, and PM2 (via MadMPI) from runtime-neutral algorithmic specifica- tions. Runtime behavior is characterized along two dimensions: performance, measured through strong-scaling experiments across computing-node counts and instances-per-node configurations visualized as 3D surfaces, and programmability, assessed through objective code metrics (cyclomatic complexity, lines of code, and function-size distribution) applied consistently across all fifteen benchmark-runtime combinations. Key findings include MASS C++’s dominance on structured spatial workloads (SugarScape: 16.2×at 96 parallel processes (ranks); superior programmability on structured workloads), PM2’s superiority on dense matrix multiplication (17×SUMMA speedup (ratio of single-process to parallel time) from explicit communication overlap), and HPX’s leading programmability on irregular workloads. This work constitutes the first algorithmically equivalent cross-runtime characterization of MASS C++, HPX, and PM2, yielding empirically grounded, workload-specific guidance for distributed C++ runtime selection.