Measuring the Capacity of a Web Server
 

 

This paper describes a new method for generating web traffic that accurately simulates real web traffic.  The paper focuses on a BSD implementation of
a web server.  The paper has insights on the shortfalls of traditional measuring methods.  The insights include the observation that traditional
measurements do not simulate the burstiness of true web traffic.  Instead they apply a constant load to web servers.  In additional traditional
measuring tools tend to be small sets of client machines that do not overload the server, and as a result remain in lockstep with the server.  In
additional LAN based test beds do not simulate WAN conditions well.  WAN conditions are characterized by bottlenecks caused by long delays, and packet
loss.

 

The architecture proposed in the paper is to have client machine running two processes.  One process is responsible for establishing connections to a
server, making a request and handing the socket of to the second process.   The second process processes the receive.  This allows the first process
to generate connections at a high rate.  

 

The experimental test bed for the paper consisted of 4 client machines, and a dual processor web server.  Test results showed that the scheme used in
this paper is capable of producing load that is not in lockstep with the server.  It also showed that using this scheme it is possible to simulate
burstiness.