Development and Tuning Framework of Master/Worker Applications

Automatic Performance Analysis

计算机科学;物理学

Anna Morajko ; Eduardo César ; Paola Caymes-Scutari ; José G. Mesa ; Genaro Costa ; Tomàs Margalef ; Joan Sorribes ; Emilio Luque

Others  :  http://drops.dagstuhl.de/opus/volltexte/2006/505/pdf/05501.CaymesScutariPaola.Paper.505.pdf
PID  :  6659

学科分类：计算机科学（综合）

来源: CEUR

【 摘 要 】

Parallel/distributed programming is a complex task thatrequires a high degree of expertise to fulfill theexpectations of high performance computation. TheMaster/Worker paradigm is one of the most commonlyused because it is easy to understand and there is a widerange of applications that match this paradigm. However,there are certain features, such as data distribution and thenumber of workers that must be tuned properly to obtainadequate performance. In most cases such features cannotbe tuned statically since they depend on the particularconditions of each execution. In this paper, we show adynamic tuning environment that is based on a theoreticalmodel of Master/Worker behavior and allows for theadaptation of such applications to the dynamic conditionsof execution. The environment includes a pattern basedapplication development framework that allows the user toconcentrate on the design phase and makes it easier toovercome performance bottlenecks.