Information Caching

Peter Danzig, University of Southern California

Donald Neal, University of Waikato, New Zealand

Stan Barber's Notes

Peter Danzig speaks:

Harvest Object Caching

A study was done to determine how much could be done by caching data on ftp. The results of this indicated that 35% of traffic could be removed by caching information.

Motivation for Web Caching

Improve perceived response time
Cut link bandwidth needs in half
Offload remote servers to obtain resilency
Avoid congestion-collapsed international links
Add robustness to the Web
Route Web requests to closest, replicated server

AOL has 4 million users and multiple hundred thousand URLs being processed by AOL ever hour.

Design

Non-blocking, multi-threaded internal architecture
DNS Cache (and negative caching)
Negative Object Cache
Order of magnitude faster than Netscape & CERN
Runs on all Unix platforms
Configurable, scalable resolution protocol
Performs expensive ops in the backround
Serves 3M requests/day/CPU

Deployment at BIG ISP

Set your browser proxy-settings
Deploy access caches for your browser
Deploy cluster of cache at access centers
Configure resolution policy to maximize locality and resiliency

HTTPD Accelerator

Adds a RAM cache to your web server
Serve 3M hits/day from desk top workstations
Support virtually-hosted httpd's
Accelerator sites on port 80 (web server on 81)

Performance of httpd accelerator (very impressive) is about an order of magniture improvement over all regular servers.

Technical Challanges

Transparency
- multiple browsers
- multilingual browsers
- different consistency models
Routing
- cache working set management
- coordinate cache routing tables across thousands of automous systems
- cache-to-cache authentication

For more information:

Questions

How is the cache updated? The consitency policy:

If there is an EXPIRED in the header, believe it.
If there isn't one, calculate something reasonable.

Matt Mathis suggests that if an ISP would do more caching services, the ISP could reduce the required size need of the ISPs transit link.

What about usage stats? It is a problem. How about adding a single non-cachable object on a page to get the accounts.

Donald Neal speaks:

New Zeland has been on the Internet since 1989 with a 9.6 baud link. PACCOM provided initial funding, but since then end-users in NZ have paid for all costs. These costs are generally allocated on a usage basis.

The cacheing service is provided as a service for which people pay. Folks get cheaper rates when they use the cache than if they don't.

There is one national cache in NZ. It is connected to two NLANR caches in the US and the Australia cache.

This not strictly a hierarchy. However, it appears that the minimum savings is about 12% and the maximum 32%. The average is 25.5%.