Resource allocation for guided parameter search applications on high-performance parallel computing environments.

摘要

Parametec study simulations, or parameter sweeps, represent an important and increasingly common class of applications which arise in many areas of engineering, including fields such as Computational Fluid Dynamics, Bioinformatics, Particle Physics, Protein Folding, etc. The efficient computation of parameter sweeps is an important challenge in computer science and has received a lot of attention.; From a computational perspective, parameter sweeps are typically structured as sets of “experiments”, each of which is executed with a distinct set of input parameters. More specifically we can define a parameter sweep as a (large) set of “independent” tasks, meaning that there is no, or little, task inter-communication.; In many cases parameter spaces can become so large that it is not feasible to compute them entirely, even on large-scale platforms. This is due both to extensive parameter value ranges and to high dimensionality of the parameter space itself. However, most users are interested in finding specific parameter space regions that match some criterion. Therefore, an appealing approach is to search the parameter space, which saves both time and compute resources. We call such applications Parameter Search Applications. When users or search algorithms are given the ability to actively interact with the ongoing search process, guiding the exploration of the parameter space, we denote such applications as Guided Parameter Search Applications (G-PSAs).; In this dissertation we focus on compute resource allocation strategies that improve the performance of G-PSAs. We propose a number of strategies and present extensive simulation experiments, which demonstrate that prioritized resource allocation can improve the performance of G-PSAs substantially. Furthermore, our study spans different application and compute platform regimes, and thus provides understanding of which resource allocation strategies are effective in which settings. This understanding represents a critical advance as it is key to enable high performance G-PSAs in practice.