Reliability-Performance-Energy Joint Modeling and Optimization for a Big Data Task

Abstract

Big data tasks has increasingly became an important kind of applications with distinctive features of large amount of data and high computational complexity. Many big data processing tools, such as Hadoop, Mesos, and Spark can split a big data task into multiple subtasks that can be executed independently. However, how to develop a rational resource scheduling strategy for running the subtasks is an important issue. In principle, parallel execution of the subtasks can improve the performance of the entire task, and redundant execution of any subtasks also has a positive effect on guaranteeing the reliability of the task, but both parallel computing and redundant computing inevitably need to occupy additional servers, which results in consuming more energy. Thus, reliability, performance, and energy factors should be fully taken into account for designing a comprehensive resource optimization strategy. In this paper, as for a big data task executed in a parallel and redundant manner, we propose a joint modeling approach to analysis important reliability-performance (R-P) and reliability-energy (RE) correlations with considering random server failures and link failures. Furthermore, a profit optimization model and a genetic algorithm (GA) searching corresponding optimal solutions are proposed for balancing the complicated performance-energy (PE) tradeoff. Illustrative examples explicitly show important R-P-E correlation, and also demonstrate that the presented optimization technique contributes to achieving a notable optimization effect on expected pure profit gained by executing the big data task.