Re: Proxy benchmark/tester

On Tue, 16 Dec 1997, Justin Pietsch wrote:

> There is an interesting paper at
> http://www.cs.rice.edu/CS/Systems/Web-measurement/ where the authors
> atempt to more correctly measure web servers. I think some of their
> methods would directly apply to attempting to simulate client latency on a
> proxy also.

Interesting paper indeed. Their _essential_ idea, as I interpreted it, is the
use of non-blocking sockets to generate more requests than a Web server can
handle. Clearly, non-blocking sockets should be used in practice for many
other reasons as well.

When suggesting a proxy benchmark/tester, I planned to use Squid-like
approach (non-blocking sockets + select loop) to generate heavy traffic with
small number of client machines. Also note that with proxy benchmarking you
probably want to simulate the server(s) as well. With your own server, you
can both reach peak loads faster and simulate "interesting" server behavior.

As for simulating client latency, I do not think that the paper addresses
this issue (they do have some suggestions on how to do it via a router
though). However, I think it is feasible to simulate "slow" or "broken"
clients to the degree we need for proxy benchmarking or testing.

Alex.
Received on Wed Dec 17 1997 - 07:28:20 MST