On Modelling and Prediction of Total CPU Usage for Applications in MapReduce Environments

TL;DR

This paper presents a polynomial regression-based model to predict total CPU usage of MapReduce jobs, aiding resource provisioning and configuration parameter selection in cloud environments.

Contribution

It introduces a novel approach to model and predict total CPU usage based on configuration parameters and input data scaling in MapReduce environments.

Findings

Prediction accuracy within 8% of actual CPU usage

Model validated on three real-world applications

Input data scaling influences total CPU usage

Abstract

Recently, businesses have started using MapReduce as a popular computation framework for processing large amount of data, such as spam detection, and different data mining tasks, in both public and private clouds. Two of the challenging questions in such environments are (1) choosing suitable values for MapReduce configuration parameters -e.g., number of mappers, number of reducers, and DFS block size-, and (2) predicting the amount of resources that a user should lease from the service provider. Currently, the tasks of both choosing configuration parameters and estimating required resources are solely the users' responsibilities. In this paper, we present an approach to provision the total CPU usage in clock cycles of jobs in MapReduce environment. For a MapReduce job, a profile of total CPU usage in clock cycles is built from the job past executions with different values of two configuration parameters e.g., number of mappers, and number of reducers. Then, a polynomial regression is used to model the relation between these configuration parameters and total CPU usage in clock cycles of the job. We also briefly study the influence of input data scaling on measured total CPU usage in clock cycles. This derived model along with the scaling result can then be used to provision the total CPU usage in clock cycles of the same jobs with different input data size. We validate the accuracy of our models using three realistic applications (WordCount, Exim MainLog parsing, and TeraSort). Results show that the predicted total CPU usage in clock cycles of generated resource provisioning options are less than 8% of the measured total CPU usage in clock cycles in our 20-node virtual Hadoop cluster.