I don't mean to be rude, but please stop this vapid series. Your last article measured HTTP throughput on a static file using localhost. I can't even begin to critique that benchmark as utterly unrealistic. And now that you've introduced a real world constraint, you find that throughput drops? That's not exactly surprising is it?

If you really want to do a realistic and robust benchmark, load up some typical database-driven apps like Wordpress, generate a sample HTTP request and response stream, snapshot the database and then benchmark the exchange by changing only the web server being used.