The impact of weights on the performance of Server Load Balancing systems

Server Load Balancing (SLB) is used when a system has to offer high throughput and High Availability (HA). Web services as offered from Amazon or Google rely on SLB with a large number of back end servers. Since these clusters of back end servers tend to be heterogeneous, Load Balancers (LBs) offer balancing algorithms like Weighted Round Robin (WRR) and Weighted Least Connection (WLC) which provide the ability to consider the different capacities of the machines by a configured weight. This paper presents benchmarks which show the influence of the chosen weight values on the performance of the SLB system and confirm simulation results done earlier in this research field. The measurements are done in a web server environment running a Wikipedia instance based on a dump from 2008 using the associated HTTP access traces. Furthermore, the measurements show that badly determined weights may lead to unpredictable effects in a SLB environment, like WRR outperforming WLC. Therefore, determining good weight values is an important task to achieve a good and stable performance. The presented benchmark results give a hint that traditional benchmarks like the Byte-Unixbench may still be a good choice to determine weights for modern web service scenarios.