We investigate the use of parameterized mi- crobenchmarks to explore Web server perfor- mance over a wider range of workloads than is tra- ditionally represented in “realistic” benchmarks. We find that these “wide-spectrum” tests not only give insights into the strengths and weaknesses of various server software, but also reveal long- standing performance problems that have either been undiscovered or improperly diagnosed. To enable our approach, we have developed “Flexi- client”, a componentized testing system that al- lows the easy creation of parameterized workload generators to test HTTP-based servers.

Even with simple static file workloads, we un- cover a number of performance problems on both popular and experimental systems, including the following: severe deterioration of FreeBSD’s min- core system call performance with growing mem- ory size on event-driven servers, process sched- uler overhead crippling persistent connection ben- efits for Apache on Linux, and low multiproces- sor performance gains on Linux. We addition- ally discover some other interesting behaviors that run counter to conventional wisdom: the most popular web server benchmark software generates surprisingly small working sets, and select-based servers may not suffer as much from idle persis- tent connections as previously assumed.