Extending the PaScal Analyzer for MPI Scalability Analysis: Design, Implementation, and Validation

Abstract

Performance analysis is fundamental in High-Performance Computing (HPC), as it is essential for optimizing both applications and the computational resources used. However, existing analysis tools are often complex or not specifically designed to assess scalability. This gap hinders the ability to identify performance trends in a practical and agile manner.

To address this gap, the PaScal Analyzer was developed—a tool for performance and scalability analysis. The program was originally designed for parallel applications on shared-memory systems. This work presents a new module for distributed applications based on MPI in Slurm environments. The initial proposal features a complete architecture using auxiliary processes to collect execution times with low intrusion. However, the currently implemented version adopts a simplified architecture that does not yet use auxiliary processes, while still preserving low-intrusion data collection.

Experimental validation was conducted using the NASA Advanced Supercomputing (NAS) Integer Sort (IS) benchmark across different problem classes and process counts, evaluating metrics such as execution time, speedup, efficiency, and scalability. The results confirm the theoretical models of Amdahl and Gustafson, demonstrating that the tool can accurately capture performance behavior in distributed systems. This facilitates informed decision-making regarding resource allocation and system scaling.