[jeffgrant]

It almost presents itself as if the web request is waiting for a thread to answer the request, in that when it does get a response, the response is fast, but there may be a substantial initial wait in a queue for that response to be handled. I'd also almost think the performance is related to available sockets. I generally tune a web server such that the TCP fin_wait and stuff is way, way shorter than usual... as in 15-20 seconds cleanup rather than the default.

This would show up as a non-contended box (cpu/mem/disk IO all being idle or nowhere near maxed) but would have **** for actual end-user performance.

But yeah, performance has been very hit and miss.