Application Development using Compositional Performance Analysis

A parallel programming archetype [Cha94, CMMM95] is an abstraction that captures the common features of a class of problems with a similar computational structure and combines them with a parallelization strategy to produce a pattern of dataflow and communication. Such abstractions are useful in application development, both as a conceptual framework and as a basis for tools and techniques. The efficiency of a parallel program can depend a great deal on how its data and tasks are decomposed and distributed. This thesis describes a simple performance evaluation methodology that includes an analytic model for predicting the performance of parallel and distributed computations developed for multicomputer machines and networked personal computers. This analytic model can be supplemented by a simulation infrastructure for application writers to use when developing parallel programs using archetypes. These performance evaluation tools were developed with the following restricted goal in mind: We require accuracy of the analytic model and simulation infrastructure only to the extent that they suggest directions for the programmer to make the appropriate optimizations. This restricted goal sacrifices some accuracy, but makes the tools simpler and easier to use. A programmer can use these tools to design programs with decomposition and distribution specialized to a given machine configuration. By instantiating a few architecture-based parameters, the model can be employed in the performance analysis of data-parallel applications, guiding process generation, communication, and mapping decisions. The model is language-independent and machine-independent; it can be applied to help programmers make decisions about performance-affecting parameters as programs are ported across architectures and languages. Furthermore, the model incorporates both platform-specific and application-specific aspects, and it allows programmers to experiment with tradeoffs better than either strictly simulation-based or purely theoretical models. In addition, the model was designed to be simple. In summary, this thesis outlines a simple method for benchmarking a parallel communication library and for using the results to model the performance of applications developed with that communication library. We use compositional performance analysis - decomposing a parallel program into its modular parts and analyzing their respective performances - to gain perspective on the performance of the whole program. This model is useful for predicting parallel program execution times for different types of program archetypes (e.g., mesh and mesh-spectral), using communication libraries built with different message-passing schemes (e.g., Fortran M and Fortran with MPI) running on different architectures (e.g., IBM SP2 and a network of Pentium personal computers)