Evaluation of High-Performance Application Escalability on Public and Private Clouds

Abstract: Recently, subsidies granted by IaaS providers (Infrastructure as a Service) for research projects in universities have facilitated the performance analysis of high-performance computing (HPC) applications. However, the performance of HPC applications relies heavily on cloud platform I/O support, especially the virtual communication network between machines. In this paper, we evaluate the scalability of the NPB-NAS benchmark suite and applications for processing spatial data in a public cloud, Azure, and a private cloud using XenServer. As a reference, we used for comparison a cluster of physical machines without virtualization. In our experiments, the application' speed-up in the public cloud was up to three times lower for CPU-bound applications, compared with a physical cluster with similar hardware. Despite the higher quality of infrastructure in the public cloud, applications also had run-time up to three times higher compared to the runtime in the physical cluster. When comparing the application' speed-up between public and private cloud, we also found that it is not only affected by the extra layer of virtualization, but also by other cloud platform characteristics, such as the use of neighboring machines, restating results of previous work. Our results indicate that although convenient, evaluating the scalability of distributed applications on virtual platforms (public or private cloud) can lead to incorrect or inaccurate interpretations regarding the scalability of some types of distributed applications.

Keywords: Virtualization; Cloud Computing; HPC; Hypervisor.

Complete monograph. Copyright © 2016. All rights reserved.

Disclaimer: Although the student carefully wrote the original abstract, and it was revised and improved, English is not him or the advisor' mother language. The original work is written in Portuguese.

Citation: Ariel Lauber de Paula Silva. Evaluation of High-Performance Application Escalability on Public and Private Clouds. Monograph. Bacharelado em Ciências da Computação. Universidade Federal de Goiás, Regional Jataí. Jataí, GO, Brasil. 2016. 51p.