Session Chair

Session Chair: Peter Honeyman
Peter Honeyman works at the Center for Information Technology Integration (CITI) in Michigan, USA and has been instrumental in several software projects, including Honey DanBer UUCP, PathAlias, MacNFS, and Disconnected AFS. His research focus is on security in distributed systems and distributed file systems for mobile computing. He is the author of dozens of journal and conference papers and serves regularly on conference program committees. Peter Honeyman is a Director of the USENIX Association and a member of AAAS and EFF.

P. Cao, J. Zhang and K. Beach
University of Wisconsin-Madison, USA

Presented by Kevin Beach

The motivation for this work stemmed from the fact that dynamic documents are on the rise. Recent traffic analysis suggests that a third of all documents carry cookies, almost a quarter contain queries, and many popular pages now carry rotating advertising banners. The problem with traditional web caches is that they treat pages as datagrams. However, due to dynamic content, they ought to be trated as objects comprising content plus code.

Kevin proposed that mobile code should be attached to web pages in the form of cache applets which would be executed at caching proxies. A prototype based on this approach has been written in Java, the language chosen for practical reasons only. Web servers mark URL as having an associated cache applet. When the URL is requested the proxy invokes the cache applet. Examples of cache applets include those which log URL accesses and that generate client specific pages based on client preferences.

The system features security mechanisms to counter illegal third party access to the web server and deal with denial-of-service. This is done by using a type safe language, station examination of applet code and resource accounting so as to enforce consumption limitations. Each web server is a security entity and applets from different servers cannot collaborate. As a by-product of the work, it was noticed that string operations and file read accesses are not performant in existing applets.

In the system presented there are some design limitations: at the proxy there are no transactional guarantees, real-time guarantees or session semantics and servers cannot query the state of the proxy. However, the scheme does achieve successful caching of dynamic pages.

A couple of points for clarification were raised. Doug Terry asked whether additional garbage collection took place when caching objects were tossed from the system. Kevin replied yes, all objects directly associated with the caching object were also tossed. In response to a question from Armando Fox, it was stated that the approach was designed to provide better distribution to users and increase the number of clients which could be serviced.

G. Pierre and M. Makpangou
INRIA, France

Presented by Guillaume Pierre

When designing a distributed Web caching system it is necessary to make a number of choices. There are conflicting optimisation criteria (access latency, cache inconsistency, network load reduction, cost) which must be traded off against each other. Standard cache configurations don't perform too badly, but they don't account for Web traffic characteristics and the actual network connectivity. Guillaume argued that custom configurations offered a better approach.

Saperlipopette! evaluates the performance of Web caching architectures based on real use. Web traffic and network QoS data are collected and passed into a simulator which uses them to evaluates the cache architecture. The talk was concluded by saying the Saperlipopette! allows the monitoring of WWW activity and evaluation of the performance of candidate caching architectures.

Two questions were put forward. Firstly, where did the name Saperlipopette! came from? Guillaume said that it didn't really mean anything and was just a French expression of surprise. The second question asked how long traces were, the period of time it took to collect them and how long it took to replay them in the simulator. The response took the form of an example: 26 million requests were captured over a period of about one month and took about an hour to replay.

A. Fox, I. Goldberg, S. D. Gribble, D. C. Lee, A. Polito and E. A. Brewer
University of California at Berkeley, USA

Presented by Steven Gribble

Steven opened with a deliberately provactive statement: that CORBA, Java and all these things try to enable middleware, but middleware is more than just that and that this talk is a CORBA free zone! The idea for this project was spawned by wanting to query the internet movies database while in a cinema, using a 3Com PalmPilot and wireless networking.

PDA's are not, and should not be, first class internet citizens. They are resource-poor devices while Web protocols and content formats are complex. Indeed, it is nearly impossible to implement the latter on the former due to their limited software capabilities. The solution is to use a level of indirection using the distributed systems principle of moving complexity away, though this time away from the impoverished end device rather than the server in the form of an active middleware proxy. Benefits of this approach are the realisation of previously impossible features, transparent functionality upgrades, delivery time optimisation of content, enhanced performance, improved stability and access to legacy systems. Another important aspect is to support refinement of the information received at the client as users ultimately know best. The fact that the proxy is now on the critical path means that it must be highly available.

While the presented approach can be used to support PDA access to many Web pages, some just can't be supported - eg: those containing very large images, Java Applets or ShockWave information. The talk was concluded by summarising that a level of indirection solves many problems.

Neil Mason questioned whether the system had been implemented on a wider range of PDAs than just the PalmPilot? Steven replied that such hadn't been done in the research environment, but that a company had built a WindowsCE version. This extra version had been implementation with the greatest of ease since only the PDA specific part of the system needed to be changed to support the new device. Another question asked whether Java Applets were really infeasible? Steven answered that they were conceptually possible but would be extremely complex to implement.

Panel

Chair: Peter Honeyman
KB: Kevin Beach
GP: Guillaume Piere
SG: Steven Gribble

Chair: Missed
Answer missed

Chair: Could you provide compression on communications between the proxy and the PDA?
SG: Possibly, but it would have to take the form of run-time encoding.

Chair: Missed
Answer missed

John Ridgeway: People adapt. The trace of web traffic will change over time as peoples behaviour will change according to network access times. Have you considered this?
GP: Good point. We are thinking of relexive, adaptable proxies in the future.

Chair: Steven, you have a lot of users?
SG: Yes, there are currently about 11,000 users.

Question: We now have client side caching and server side caching. How do we mediate these and stop them both taking time to do the same job?
KB: I don't think that's been addressed so far.
Chair: I disagree - I've seen some papers on this topic. Web browsers are built with caching included: I'm surprised there's interest in server side caching. Also, I'd think it wasn't going to get the hit-rates and papers seem to confirm this. Server side caches are just delay servers between the client and original server.
SG: There is a counter-example to this: the recent publication of the Starr report created meltdown!

Comment: Computer scientists are not currently building systems based on cost models. These caches should be being used to reduced costs over expensive links.
GP: It's hard to envisage a global caching system.
Chair: Couldn't you just add a cost model to your system?
GP: Perhaps.

Question: Any thoughts about server operators who want to subvert caches to improve their hit rates and sponsorship income?
KB: That's why we have special applets for logging and advertisements.

Chair: Missed
Answer missed

Question: Getting at the content of Web documents is hard. Can it always be automated? Should we not adjust the standards?
SG: SGML work is moving towards that.