W3C Activities Propagation

Replication and Caching Position Statement

Since the Web started in 1990 is has shown an exponential growth, and now - five years later - the growth is still exponential. No one knows how or when the growth will slow.

The simplicity of the Web model was one of the main reasons it succeeded. The Web model is based on URIs, HTML, and HTTP. However, HTTP by itself, as it currently exists cannot meet the robustness or scaling requirements that are clearly needed due to the Web's success. HTTP does not provide for recovery in the face of network problems, and often information is unavailable due to single point failures.

There is an urgent need for making the Web more mature in order to scale to a number of at least 100 times the current size and robust enough for mission and life critical applications. Network hot spots (or "flash crowds") often appear in the internet when particular information has large topical interest, causing major problems both to end users (poor response) and denial of service to users of the network near the hot spot (due to the resulting congestion in nearby parts of the Internet). Efficient techniques for replication and caching is a corner stone in achieving the goals of scale and robustness as well as responsiveness; how much bandwidth a good caching and replication system could save is currently an open research question.

Problem Statement

There are currently a number of distribution protocols on the internet which are suited to different patterns of readership. Please have a look at the problem statement written by Tim Berners-Lee. The protocols differ in the conditions under which information is transferred:

SMTP (Simple Mail Transfer Protocol)
Transfer on author demand. Readership defined at time of publication, in 1..10,000 range.
NNTP (Network News Transport Protocol)
Transfer unconditionally to all sites. Wastes resources globally when used for small readership, and doesn't scale very high. The replication of individual news groups relies on social bindings between system administrators of News servers
HTTP (Hypertext Transport Protocol)
Transfer on receiver demand. Readership unlimited. Network suffers when readership is high.

When an individual makes a piece of information available, currently that individual's chooses the application, which implies the protocol which will be used. Misuse of resources or bad performance occurs when the wrong protocol is used. Here are three examples of a use of one protocol when another would have saved resources.

Recent studies by Margo Seltzer shows that flash-crowds are very common. Documents on a single server generally do show good locality. That is - popular documents are really popular! Other studies have shown that documents often are transmitted thousands of times between changes to the document; a system which cached popular documents where the caches are (conceptually) on the other end of a communications link should show good hit rates, rather than just at the periphery of the network. Research is underway to prove/disprove this hypothesis.

The Web is moving towards a more dynamic content. Web applications are becoming customized on a per user basis. This has already a significant impact on the efficiency of caching throughout the Web. Dynamic content can be expected to continue to grow; however, it is possible to break down the content in smaller static parts which in terms are cachable. Java and similar mobile code systems enable the building of Web applications in which the content as presented to the user may be highly customized, while most of the program and data required to perform the customization may remain cachable. The long term effect of dynamic content on cachability of information is therefore uncertain; current trends have been reducing cachability, but the longer term outlook is possibly more favorable than the current trend. Study of Web traffic is vital to understand the trends of Web traffic.

Many content providers need to have some way of determining how much their information is being used. While providing useful or interesting content may be supported by advertising, even people providing information for free need basic information to support their activities. People who are marketing products, or providing public information (e.g. government agencies) need to be able to justify their expenditures. Due to inadequacies in HTTP/1.0, information providers have often disabled caching (with bad results to the users) just to get elementary usage statistics. Mechanisms such as the hit count proposal of Jeff Mogul (Digital) and Paul Leach (Microsoft) being discussed in the IETF HTTP working group in combination with HTTP/1.1 facilities may enable more of the Web content to be cachable, though content providers desiring full demographic information may still generate significant network load.

The Web provides an escape route for introducing efficient caching systems by its support for proxies. The CERN server experienced an enormous demand as a result of being the first proxy server. Current proxy servers are placed at the leaves of the distribution tree a long way from the origin server. As a result, most proxy servers seldom achieve a cache hit rate of more than 50%.



HTTP/1.0's caching facilities are poor and error prone. Application builders have often had to disable caching (at the expense of performance to the end user) . Getting the slightly stale answer quickly may be desirable for many applications, but some applications (e.g. electronic commerce) need to get absolutely correct answers. One of HTTP/1.1's major design goals is to allow reliable caching of content, and applications to work reliably in the face of caching proxies. A proxy network must be able to provide the level of consistency that the application developer has specified.


One could define the efficiency of the system may as the ratio between the reference network load and the actual network load, where the reference network load is the sum over network hops of bytes transferred using simple HTTP with no proxy network. The actual load is the cost in actual byte/hop products, which in a caching system has to be averaged over all connected access of the source document.

Response time

A fundamental requirement of the Web is that access time for a random document should be low, whatever the reader's reading history to date may have been. Studies indicate human problem solving is impaired when retrieval times rise above 100ms. Global network latencies are easily of this order. Preloading of documents likely to be fetched may be needed to meet response time constraints; but excessive bandwidth use for documents not used mitigates against blind use of this technique.

In practice in the Wide World, as bandwidth and CPU speeds rise, the number and length of round-trip delays becomes the limiting factor. The response time requirement translates into a protocol requirement for minimizing the number of serial transactions with different parties, and the number and length of round trip times for each.

Storage Limitations

We must assume that any participating node in the system has a limited amount of storage to make available for caching objects and metainformation involved in the protocol. As this storage reaches capacity, the system should degrade gracefully.

There is an overall web requirement for scalability. Given finite storage at each node, scalability rules out systems in which complete network maps or hypertext maps or name tables are kept at any one node.

Limited Human Intervention

A human must dedicate resources (storage, processing, bandwidth) to the use of the system at some point. Experience with successful systems in the global Internet has shown that end users or system administrators should not have to be involved in making routing or optimization configurations in large scale systems; not only is such configuration expensive in human time, such configurations become obsolete, and systems fail as the network evolves if the system is not dynamic in nature. For a large scale system, most replication and caching decisions must be completely automatic, though local policy decisions may be imposed over such an infrastructure.

The Essence of the Caching Problem

There are two complex problems involved.

One is the multicast routing problem, which is the generation of the distribution tree for optimum efficiency for a given readership pattern (known or expected) and a given network topology.

The second is the name resolution problem. When a client first requires access to an object, for which a multicast route may already exist, it must in a short time, find the location of the nearest copy. Possible solutions might include an initial query to the original server, so long as in future access can be directed to a closer copy if available. Most cache systems involve grouping objects, and assuming a correlation between accesses to documents in the same group. HTTP's basic rule is that the URL of a document, once the initial domain name has been resolved, is a random string to which no semantics may be attached. There is a provision for the use of hierarchical spaces (used with relative addressing) but there are no numerical constraints on the number of objects at the same points in a hierarchy: indeed, there are webs containing an infinite number of virtual documents.

In the future we expect URIs to become more like names and less like addresses. This means that any control information which enables cache efficiency may not be encoded in the URIs, and so must be transferred in metadata, essentially in list of names. The less structure there is in a name, the more difficult it is to scale the storage of the name tables; the more structure there is, the more likely it is that the name will later have to be changed as the storage structure changes. This is an essential problem of naming.

By combining information of both object access and link traversal, an information source may be able to provide information to a replication/caching system to improve user performance, by preloading of documents likely to be accessed next into downstream caches. The topology of the web therefore, as well as of the internet, may be useful to such a system.

Any solution is likely to be sound engineering compromises combined with some good algorithms.. This is an area for directed research, and grass roots experimentation, and it will be interesting to see which produces the results most effectively.

Next Steps

Short Term

The new caching facilities of HTTP/1.1 need full testing before HTTP/1.1 becomes a full Internet standard. We believe the new facilities of HTTP/1.1 allow robust applications to be built in the face of caching. This belief needs proving, and our Jigsaw server is one of the first servers providing both basic HTTP/1.1 and proxy support. Full validation of the HTTP/1.1 design needs to be complete to enable graceful evolution of the Web, so that future caching and replication systems can be deployed incrementally via the proxy mechanisms built into all Web software.

We have also implemented HTTP/1.1 in both our library (libwww) and the Jigsaw HTTP server, now available for anyone to use for testing of HTTP/1.1 implementations. W3C is fostering a mailing list for implementors of HTTP/1.1 to exchange implementation experience and solutions to implementation problems. A few remaining small HTTP/1.1 design problems are currently being discussed on the IETF HTTP working group mailing list.

We believe HTTP/1.1 fixes HTTP's worst bugs and inefficiencies. Having completed this work, we should spend some time to understand to what degree further compression and/or reengineering of HTTP itself would actually be beneficial versus work on the general caching/replication problem.

Work in the distributed authoring and versioning area must be tracked to understand the implications to caching and replication of content.

Medium Term

Many major sites today (e.g. AltaVista, Netscape, Microsoft) are now replicated on a world wide basis, generally using a combination of ad-hoc facilities (e.g. AFS, DCE DFS, rdist, rcp, tapes delivered via overnight delivery, etc.). For example, AltaVista is replicated on three continents (North America, Australia, Europe), and each replica may actually consist of a set of replicas. We can expect this trend to continue for successful sites to handle the scaling load.

However, the end user has to know which replica to use, and in the case of failure or network difficulties, the user may still find the replica they use unavailable. Naive users, now by far the major users of the Web, are often unable to work around problems when they arise, and discovering which replica is "closest" to a user is left to the users, who has little information to guide their use.

Therefore, facilities to allow automated, transparent selection of the "closest" available replica of a service without user intervention would go a long way toward reducing long haul or intercontinental network load (by far the most expensive), while providing end users with much higher performance, and much better robustness. Whether these facilities will be built into end user browsers or into ISP's proxy caches, and what exactly these algorithms might be is currently an urgent, open question.

The Internet infrastructure has much better knowledge as to what site may be "closest" to a user; however, there has been little communication to date between developers of the Web and developers of the Internet routing infrastructure. To solve this problem we believe will require cross fertilization between many parts of the IETF (Internet Engineering Task Force) and the Web community. We plan to spend significant effort fostering such ties, and evaluate possible solutions that may appear.

While W3C might prototype possible solutions to this problem, it is as likely or more likely that a solution to this problem will come from elsewhere, as the expertise to help with solving the "closest copy" problem is in the general network community rather than the Web community.

Long Term

Over the last 20 years a large number of different techniques have been developed for handling caching and replication in a distributed environment. Such techniques can be applied to the Web. It is clear that not one single technology will solve the problem but some of the key technologies involved are:

W3C sees it as an essential part to bring together research and industry in order to create a shared understanding of the problems facing the Web. In April 1996, W3C organized a workshop on Web robustness involving many of the inventors of the Internet and key representatives from industry. Fostering further interactions and experimentation with prototypes (e.g. Squid) systems need to continue, until one or more solutions are found.

We need answers to basic questions, for example:

These a just a few of the open questions. The W3C will foster research on these problems and experimental deployment of such systems.

Jim Gettys, Tim BL, and Henrik Frystyk Nielsen,
@(#) $Id: Activity.html,v 1.8 1997/08/09 17:54:50 fillault Exp $