User-Centric Heterogeneity-Aware MapReduce Job Provisioning in the Public Cloud

Cloud datacenters are becoming increasingly heterogeneous with respect to the hardware on which virtual machine (VM) instances are hosted. As a result, ostensibly identical instances in the cloud show significant performance variability depending on the physical machines that host them. In our case study on Amazon’s EC2 public cloud, we observe that the average execution time of Hadoop MapReduce jobs vary by up to 30% in spite of using identical VM instances for the Hadoop cluster. In this paper, we propose and develop U-CHAMPION, a user-centric middleware that automates job provisioning and configuration of the Hadoop MapReduce framework in a public cloud to improve job performance and reduce the cost of leasing VM instances. It addresses the unique challenges of hardware heterogeneity-aware job provisioning in the public cloud through a novel selective-instance-reacquisition technique. It applies a collaborative filtering technique based on UV Decomposition for online estimation of ad-hoc job execution time. We have implemented U-CHAMPION on Amazon EC2 and compared it with a representative automated MapReduce job provisioning system. Experimental results with the PUMA benchmarks show that U-CHAMPION improves MapReduce job performance and reduces the cost of leasing VM instances by as much as 21%.