Using High-Performance Computing Profilers to Understand the Performance of Graph Algorithms
Costain Nachuma · ScholarWorks @ The University of New Orleans (The University of New Orleans) · 2020
An algorithm designer working with parallel computing systems should know how the characteristics of their implemented algorithm affects various performance aspects of their parallel program. It would be beneficial to these designers if each algorithm came with a specific set of standards that identified which algorithms worked better for a specified system. Therefore, the goal of this paper is to take implementations of four graphing algorithms, extract their features such as memory consumption, scalability using profilers (Vtunes /Tau) to determine which algorithms work to their fullest potential in one of the three systems: GPU, shared memory system, or distributed memory system. The features extracted in this study were scalability, speedup, and parallel efficiency. We find that when looking at various parallel algorithms: Community Detection, Communities through Directed Affiliations (Coda), BigClam, and Breadth First Search all achieved noticeable speedup with increasing number of cores.