Effective Open-Source Performance Analysis Tools

Introduction

Performance optimization of computational software is often an iterative and tedious process. The algorithms developed by the scientist/engineer may work well on one computer architecture and may need further performance tuning on another architecture. The need to compare performance characteristics across many hardware architectures is a routine task performed when benchmarking high performance computing systems for procurement. Open-source tools for performance analysis are suited to compare application performance characteristics across a wide range of computing architectures. The main goal of this chapter is to outline a few open-source development and performance analysis tools that the author found to be effective in a wide range of computing hardware.

Today, CPUs using the x86 or x86_64, power instruction set architectures are widely used for scientific high performance computing. Single core processors gave way to dual and quad core processors mainly due to power and thermal constraints. By 2011, six core to twelve core processors from Intel and AMD are expected to be widely used in HPC systems. Another clear trend in scientific computing hardware is the use of general purpose graphics processing units (GPGPUs) as accelerators for HPC applications. Programmed using high level programming languages like CUDA (Nvidia Corporation) and OpenCL implementations (Khronos group), these commodity products are redefining the architecture of HPC systems especially at the low and medium scale deployments. The capability supercomputing space is dominated by massively parallel distributed memory supercomputers. These machines often use multi-core processors running at relatively lower frequencies enabling higher packaging densities.

The most widely used programming model for distributed memory parallel computers is based on the Message Passing Interface (MPI). Most recent distributed memory machines use multi-core chips and therefore shared memory programming models like OpenMP (OpenMP ARB) is used to exploit parallelism at the node level. This helps applications scale better by reducing the inter-processor communication through the interconnect (Smith L & Bull M, 2000). As the number of processor cores in a single chip increases, contention for shared resources within a single multi-core processor becomes a major issue preventing linear scaling. Recent processor designs focus on techniques to reduce contention for shared processor resources like memory bandwidth. The use of accelerators in HPC systems introduces an additional layer of software and hardware components. Effective analysis of the performance characteristics of software on these hybrid systems is an active area of research.

Analysis and understanding of performance bottlenecks on scientific computing applications across systems of varying hardware architectures is an important task for HPC practitioners. Standards based open source performance analysis applications facilitate the comparison of application characteristics across many hardware architectures. They are available on a wide range of systems and often offer an independent tool-chain in addition to those offered by the vendor of the HPC system.