SP: Collaborative Research: Rapid Evolution of Transport Protocols

点击翻译按钮获取中文摘要

The Internet is a tremendous engineering success. It has sustained a heady rate of growth and changeover at least the past decade and has found broad application that has made it part of the fabric of society.Driven by changes in the underlying technology and by new applications, it is certain to continue evolving.Such evolution is supported by a large and diverse body of research that often anticipates future needs.In the case of transport protocols, our focus, the future holds protocols for a new generation of streamingmedia, stable congestion control at very high rates, mobile, multi-homed and aggregate connections, andinformation dissemination via peer-to-peer networks and application-level multicast-at least.Unfortunately, it is very difficult to move from the results of such research to use in the global Internet,no matter how promising the research. Today, operating system vendors, not researchers nor even productdevelopers, must typically provide new transport protocols. To be suitable for deployment, these protocolsmust be able to successfully interoperate with the installed base of transports during the long period ofgradual adoption. This process is costly and time-consuming. More insidiously, it has an incentive structurethat discourages experimentation and early adopters. The bottom line is that current upgrade mechanismsare arguably the limiting factor in realizing the benefits of many ongoing research efforts.The researcherss propose to investigate a new approach to the evolution of transport protocols that will allow experimenters to quickly and easily upgrade transport protocols. It will do so in a realistic Internet setting and with less concern for the installed base. This would allow, for example, the GRID community to deploya TCP carefully tuned for high-rate transfers, or researchers studying the Congestion Manager to readilyexperiment with it at a large scale and across the wide-area. The system envisioned will fundamentallydiffer from today's independent end-system upgrades by allowing one end of a connection to upgrade theother by using untrusted mobile code. This "dual-end" upgrade model will encourage the deployment ofuseful protocols by tipping the balance of the incentive structure in favor of early adopters. It will encourageexperimentation by allowing researchers and developers to upgrade transport protocols directly, rather thanindirectly via operating system vendors. The proposers believe it has the potential to multiply the benefits of research that is currently being undertaken by others.Intellectual Merit. The key challenge is to determine how it is possible to realize the straightforwardvision of rapid evolution via dual-end upgrades without negating its practical benefits. This is a difficulttask, as the use of untrusted mobile code has been previously explored in the much more general domain ofactive networks, with limited success. However, they argue in the body of this proposal that now is the timefor a concerted research effort on transport protocol deployment. Recent advances in the understanding ofnetwork protocols (specifically TCP friendliness and robust congestion signaling with ECN nonces), understanding of the sharing and termination issues in safe-language runtimes, and maturing language technology (in the form of Cyclone, a safe yet C-like language)-plus a well-defined problem domain that simplifies protection issues-have made the effective use of untrusted mobile code a real, though challenging, possi-bility. The PIs have expertise in all of these areas, along with an established research collaboration.Broader Impacts. The proposed research has the potential to achieve a broad impact by a true multipliereffect: it directly helps to realize the benefits of research that is undertaken by others, encouragingrather than discouraging upgrades by virtue of its incentive structure. The ease of experimenting with newtransport protocols in a realistic setting will lower the research barrier, helping a larger pool of researchers (and particularly minority institutions) to compete more on the basis of their ideas and less on the level of deployment resources they are able to muster. The researchers expect there to be a strong educational component of this work, since students as well as researchers can safely experiment without concern for adversely impacting the network, even on generic, shared machines. Real deployment, perhaps solving real problems, can occur among fellow students, and is ultimately limited only by a protocol's value and the students' ability to publicize it. Thus will students, and researchers, be empowered and motivated.