Shadura O. Stochastic optimization of the performance of a particle transport simulation package in High Energy Physics

The dissertation is dedicated to study the optimization strategies using genetic algorithms to achieve the best computing performance of the particle transport simulation package GeantV for the High Energy Physics simulations. It had been developed to solve the problems of processing the large data sets in experiments at the Large Hadron Collider (LHC). In the scope of this dissertation had been developed a software library integrating the genetic algorithms as a part of the GeantV software package. It will be used to optimize the performance of GeantV software package for the particle transport simulation, when used for processing the physics data, collected by experiments on the LHC. The mathematical model of the non-centered principal components analysis method is defined together with the formula for estimation of error approximation. The estimation of the error approximation shows that the procedure of reducing the data matrix, based on the selection of eigenvectors of the matrix of non-centered second moments for which it has the smallest eigenvalues, is correct. A modification of the genetic algorithm is defined by introducing into the standard set of genetic algorithm operators (selection, mutation, crossover), a new operator, which is determined by the non-centered principal component analysis method (as a new genetic operator to be used on genetic populations). The results of this study are presenting the proof of the concept of optimization of GeantV performance using evolutionary algorithms with an average gain of 20% of over not optimized run in heterogeneous computing environment. The same method can be used to deploy GeantV applications on supercomputers and clusters for the efficient high-performance computing, while configuring massively parallel GeantV simulations, launched in a non-homogeneous computing environment and providing optimal scalability in high-performance computing environment. Keywords: genetic algorithm, dynamic system, fixed points, stochastic optimization, genetic operator, principle component analysis, performance optimization.