An increasing amount of information is currently becoming available th
rough World Wide Web servers. Document requests to popular Web servers
arrive every few tens of milliseconds at peak rate. To reduce the ove
rhead imposed by frequent document requests, we propose the notion of
caching a World Wide Web server's documents in its main memory (which
we call Main Memory Web Caching). We show that even a small amount of
main memory (512 Kbytes) that is used as a document cache, is enough t
o hold more than 60% of the documents requested. We also show that tra
ditional file system cache management methods are inappropriate for ma
naging Main Memory Web caches, and may result in poor performance. Bas
ed on trace-driven simulations of several server traces we quantify ou
r claims, and propose a new cache management that dynamically adjusts
itself to the clients' request pattern and cache size. We show that ou
r policy is robust over a variety of parameters and results is better
overall performance.