The World Wide Web is now a critical communication infrastructure for society. However, the foundations on which the Web is built are largely ad hoc and haphazard, with a protocol framework consisting of the hypertext transport protocol (HTTP) layered on top of the Internet's universal reliable delivery protocol-the transmission control protocol (TCP). Consequently, the performance of HTTP over TCP suffers from a number of deficiencies that are all rooted in the fact that TCP's byte-stream semantics are often at odds with user goals and the performance capabilities of end clients and servers.
Several ad hoc remedies have been proposed earlier to compensate for these deficiencies, e.g. parallel connections, persistent connections, and ACK pacing. However, all these solutions are overly constrained by the presumption of backward compatibility. We believe that the time has come to adopt a holistic and integrated approach embodied in a new transport protocol tailored specifically for the Web. Herein, we propose a set of research tasks to study, design, implement, test and evaluate our new ``Web Transport Protocol" or WebTP.
Our approach is based on a conceptual model of user validation, application semantics, network and client states, and a flexible protocol that reflects this syntax and semantics. The conceptual model emphasizes the interaction between document structure, network state, and user preferences. Our research prototype will leverage the well-established and highly successful congestion control framework of TCP and vastly improve upon the existingWeb protocols by bringing user requirements into the design loop. To do so, we will apply the principles of Application Level Framing (ALF) to the design of WebTP. ALF says that to optimize communication paths within the application and across the network, an application's semantics (i.e. its user requirements) should be explicitly reflected in the design of its network protocol. Although ALF has been adopted in modern protocols like the Real-time transport protocol (RTP), paradoxically it has not been systematically applied to previously existing applications like the Web.
We will study the following questions:
* How do we best structure the ALF-based Web protocol in terms of congestion control algorithms, loss recovery schemes, packet framing methodology, header formats, toolkit APIs, and so forth? How do these design decisions impact network performance, user satisfaction, protocol complexity, and server load?
* How then does the client elicit user preferences in the least obtrusive fashion?
* How does the receiver (client) learn of the content structure of a server's Web data and combine that structure with its estimate of network state and user preferences to control selective transmission via WebTP?
* How should functionality and state information be distributed among the source (server or proxy), client, and user and what are the design and performance tradeoffs? How does WebTP scale with increased speed and faster clients?
Our approach to these questions draws upon network engineering and communication theory, signal processing and stochastic control. We will use existing network emulation infrastructure from the the INDEX Project to comprehensively test and evaluate WebTP. This project holds the potential to dramatically impact the way we think about Web transport and will further contribute a design methodology, an actual protocol architecture, and performance analysis and results. We believe that our versions of WebTP will demonstrate dramatic improvements and pave the way for further developments.
