Web Proxy Caching: The Devil in in the Details

R. Caceres, F. Douglis, A. Feldmann, G. Glass, M. Rabinovich (AT&T Labs), 199?

Summary. By taking into account low-level details ignored in previous studies (aborted transfers, cookies, connection setup time, slow start) the simulated benefit of web caching is less pronounced than previously published:

hit ratio: ignored cookies: 54.5% hit ratio; with cookies: 35.2%.
bandwidth savings: well-connected clients should just abort; non-well-connected clients might consider continuing the download since most of the document might be downloaded anyway, allowing subsequent requests to hit in the cache.
user-perceived latency: showed 8% improvement. Connection cache grew improvement to 28%.

More Detail...

Proxy caching attempts to improve performance in three ways:

reduced user latency
reduced network traffic
reduced server load

Plus: could decrease network costs for ISP.

Two common measurements:

hit ratio
byte hit ratio (hit ratio in terms of number of bytes)

These may estimate the reduction in bandwidth but do not address the other potential improvements.

Previous studies ignored the effect of cookies (methods of customizing resources on a per-user basis; as a side effect, documents with cookies not cacheable), aborted transactions (which can increase bandwidth). This study takes into account cookies, aborted transactions, the network environment (bandwidth mismatches), slow-start (affects latency).

Build a web simulator called PROXIM. Used a trace from the AT&T Worldnet ISP.

Key performance findings above in summary. More detail on bandwidth usage:

bandwidth went *up* 18% (simulation) if the proxy continued to download a page after an abort;
if download aborted immediately, the bandwidth depends upon the magnitude of the bandwidth mismatch from the client-proxy to proxy-internet.
- 45Mbps internet connection: 8% increase in bw;
- 1.5Mbps : 2% increase;
- .5Mbps (T1) : 15% savings.

Criticism

HTTP 1.1 addresses cookies and caching. How? When will most be using 1.1? Will hit ratios then go up to the Gribble & Brewer 60% measurement?