Performance designed architectures: A neural network approach

ABSTRACTThe requirements of new applications always motivate the development of new systems. Theenormous performance range offered by today?s systems adds a level of difficulty toperformance evaluation methods, which must consider the relative values and contributions ofvarious components. It is complicated to predict and validate the contributions of theseinterrelated factors only analytically. Benchmarking is a popular way but equally open tomisuse. There are few accurate performance analysis and visualization tools, however theyare usually either too complex or system dependent.PACE is one of the easy to use and reliable performance analysis toolsets. It allows users tomodel their own system and application easily and gives an accurate prediction of executiontime. Modeling user specific systems or adapting user?s own codes into PACE still involves alevel of effort.This thesis investigates the possibility of predicting performance of real applications by usingartificial neural network models. Neural networks can learn to approximate any function andbehave like associative memories by using just example data that is representative of thedesired task. The models presented here are aimed to be simple and usable in general cases.The contribution of this thesis is to present two neural network models for performanceprediction. These models can be incorporated into a characterization tool, such as PACE, orcan be used separately. They are potentially robust and the prediction results are proved to beaccurate. Although, the artificial neural networks have been used for various prediction tasks,their use in performance evaluation area is a novel approach.Keywords: Neural networks; Performance analysis; Performance evaluation; Parallelcomputers; Parallel computingxiii