Another thing to consider beyond simply the fact that your particular benchmark ran into a network bandwidth bottleneck: web server benchmarks should not be conducted using ApacheBench until Apache makes AB multi-threaded.

Use a multithreaded benchmark tool such as WeigHTTP ( http://redmine.lighttpd.net/projects/weighttp/wiki ). WeigHTTP is essentially identical in behavior to ApacheBench, but with a -t argument to specify the number of threads.

You can approximate the same behavior in ApacheBench by kicking off multiple AB instances in parallel, but then it is up to you to aggregate the results.