Dizertačné práce

Modelovanie a optimalizácia výkonnosti paralelných algoritmov

Autor práce: Ing. Filip Janovič
Školiteľ: Prof. Ing. Ivan Hanuliak, CSc.
Dátum obhajoby: 21.8.2013
Študijný program: 9.2.9 Aplikovaná informatika
Oponent 1: prof. Ing. Petr Dostál, CSc.
Oponent 2: prof. Ing. Imrich Rukovanský, CSc.
Oponent 3: doc. Ing. Ladislav Schwartz,PhD.

Anglický abstrakt:
The theses “Performance modelling of parallel algorithms“deals with a problematic of complexity and performance evaluation of matrix parallel algorithms (MPA). The main goals were to perform • analysis of the development trends in the world, their actual state and perspectives in the areas parallel architectures, parallel algorithms, criterions for performance evaluation and computing parallel models • mathematically supported formulations of complex performance modelling of MPA including their most important components (computation, communication control and synchronisation) • theoretical and practical analysis on all used forms of parallel computers (NOW, SMP and Grid) of MPA including decomposition models (blocks, strips) analytical performance relation ♦ computation latency ♦ speedQup ♦ effectivity performance prediction ♦ issoefficiency ♦ optimisation • methodology of suggestion and its verification to measure MPA of single process parallel proces in a shared memory parallel proces in a distributed memory. For real communication models we have been demonstrated the role of decomposition model on the practical examples of parallel multiplication and the way of deriving technical  25 parameters of Gauss elimination method (GEM). For these purposes we have developed following parallel algorithms • parallel multiplication standard decomposition model alternative decomposition model • GEM shared memory (OpenMP) distributed memory (MPI). For these parallel algorithms we have derived and verified analytical equations on various practical simulation experiments. The achieved and verified results could be methodically applied to other matrix parallel algorithms for actually dominant parallel computers based on SMP, NOW and Grid. KeyWwords: parallel computer, parallel algorithm, performance evaluation methods, decomposition model, communication overheads, performance modelling, speedQ up, efficiency, issoeficiency, NOW, Grid, SMP, scalability, shared memory, distributed memory, Gauss elimination method, OpenMP, MPI.

Autoreferát dizertačnej práce


Partneri FRI

Projekty FRI