Predicting performance and scaling behaviour in a data center with multiple application servers

摘要

As web pages become more user friendly and interactive we see that objectssuch as pictures, media files, cgi scripts and databases are more frequentlyused. This development causes increased stress on the servers due to intensifiedcpu usage and a growing need for bandwidth to serve the content. At thesame time users expect low latency and high availability. This dilemma canbe solved by implementing load balancing between servers serving content tothe clients. Load balancing can provide high availability through redundantserver solutions, and reduce latency by dividing load.This paper describes a comparative study of different load balancing algorithmsused to distribute packets among a set of equal web servers servingHTTP content. For packet redirection, a Nortel Application Switch 2208 willbe used, and the servers will be hosted on 6 IBM bladeservers. We will comparethree different algorithms: Round Robin, Least Connected and ResponseTime. We will look at properties such as response time, traffic intensity andtype. How will these algorithms perform when these variables change withtime. If we can find correlations between traffic intensity and efficiency of thealgorithms, we might be able to deduce a theoretical suggestion on how tocreate an adaptive load balancing scheme that uses current traffic intensity toselect the appropriate algorithm. We will also see how classical queueing algorithmscan be used to calculate expected response times, and whether thesenumbers conform to the experimental results. Our results indicate that thereare measurable differences between load balancing algorithms. We also foundthe performance of our servers to outperform the queueing models in most ofthe scenarios.