A Caching Relay for the WWW

Steve Glassman, DEC SRC

One-line summary: A caching proxy with a 30-50% hit rate; size-limited cache, with aging heuristic used to evict documents. Users don't notice when it's turned off.

Overview/Main Points

Zipf distribution used to model access pattern: n'th element accessed with probability 1/n
30-50% hit rate for 80MB cache
Users don't notice when it's turned off: they're used to wide variance in WWW latencies (latency to go outside SRC firewall is 6-9 sec without cache)
Heuristic: check last mod time of a page, and assume it won't change for at least that long.

Relevance

Caching is needed on the Web. This study suggests a single cache is not very useful compared to cooperative caching, which isn't addressed.

Flaws

Histogram of workload (distrib. of sizes of cached objects) would be nice.
No indication of how effective is the aging heuristic--but it does highlight the fact that some such mechanism is needed--even worse if we try to do semi-coherent cooperative caching.
Is Zipf distribution really a good model for accessed elements? Need a "Web survey" of some kind to back this up.

Back to index