Adaptive orchestration of resources in distributed wide area large scale infrastructures

Abstract

The goal of this thesis is to take a step to understanding the resource management of massively and scattered distributed systems. A framework to predict grid resources behavior and leverage the execution of long running tasks over computational grids has been developed. This framework employs statistical analysis for estimating the resource behavior and uses a divisible load approach to increase the throughput and reduce the idle time exhibited by computational resources. The proposed approach focuses on an opportunistic pull resource selection mechanism: a number of very light agents are deployed in nonintrusive way running in a user space. Initially the framework collects information on user requirements and application deployment, assigns a subset of jobs to available resources, and periodically the selected pool of resources is updated to opportunistically choose the resources that better complete the assigned jobs. The statistical analysis process evaluates in run time different probabilistic functions to determine the one that better model a resource behavior. Experimental results show a significant reduction of the application makespan along with good estimations of the resource behavior.