Parametric micro-level performance models for parallel computing and parallel implementation of hydrostatic MM5

This dissertation presents Parametric micro-level performance models and Parallel implementation of the hydrostatic version of MM5;Parametric micro-level (PM) performance models are introduced to address the important issue of how to realistically model parallel performance. These models can be used to predict execution times and identify performance bottlenecks. The accurate prediction and analysis of execution times is achieved by incorporating precise details of interprocessor communication, memory operations, auxiliary instructions, and effects of communication and computation schedules. The parameters provide the flexibility to study various algorithmic and architectural issues. The development and verification process, parameters and the scope of applicability of these models are discussed. A coherent view of performance is obtained from the execution profiles generated by PM models. The models are targeted at a large class numerical algorithms commonly implemented on both SIMD and MIMD machines. Specific models are presented for matrix multiplication, LU decomposition, and FFT on a 2-D processor array with distributed memory. A case study includes comparison of parallel machines and parallel algorithms. In a comparison of parallel machines, PM models are used to analyze execution times so as to relate the performance to architectural attributes of a machine. In a comparison of parallel algorithms, PM models are used to study performance of two LU decomposition algorithms: non-blocked and blocked. Two algorithms are compared to identify the tradeoffs between them. This analysis is useful to determine an optimum block size for the blocked algorithm. The case study is done on MasPar MP-1 and MP-2 machines;The dissertation also describes the parallel implementation of the hydrostatic version of MM5 (the fifth generation of Mesoscale Model), which has been widely used for climate studies. The model was parallelized in machine-independent manner using the Runtime System Library (RSL), a runtime library for handling message-passing and index transformation. The dissertation discusses validation of the parallel implementation of MM5 using field data and presents performance results. The parallel model was tested on the IBM SP1, a distributed memory parallel computer