The Protocol Bypass Concept for High Speed OSI Data Transfer C. M. Woodside*, K. Ravindran** and R. G. Franks* * Telecommunications Research Institute of Ontario, Dept. of Systems and Computer Engineering, Carleton University, Ottawa, Canada K1S 5B6 ** Dept. of Computer and Information Sciences, Kansas State University, Manhattan KS 66506 Abstract A protocol bypass is a fast processing path which is used for some data units, for instance for large data packets. It can provide conformance with standardized layered protocol system like OSI, together with some of the performance benefits of a “lightweight” pro- tocol. The concept is discussed as it applies to the movement of user data by the OSI transport and session protocols, with some implementation experience, and an outline for an approach to formally deriving bypass specifications from protocol specifications is given. This outline uses some steps which must still be proven to be practical. Correct interleaving of data units from the two paths is a major concern, especially with mul- tiple asynchronously specified layers. It seems that the difficulty can be overcome and the concept has promise. In the (rather limited) implementation, bypassing consistently outperformed parallel processing as a means of performance enhancement. 1 Introduction The search for generality, flexibility and standardization in communications protocols has led to the OSI layered system [1],[2] and many offshoots such as MAP [3]. However, the slowness of execution of the protocol implementations, which is essentially due to the complex checking of conditions that is done at every operation, is becoming a limiting factor in some applications. Therefore a new generation of lightweight or high-speed data transfer protocols is now emerging. The lightweight protocols exploit the low error rates of many networks, and use larger packets, reduced options to unclutter the data path, and more efficient methods for congestion control. They are partly motivated by the high speeds of new fiber-based networks, partly by high-throughput applications such as file system backups and full motion video, and partly by performance constraints already being felt with current protocols. Examples are Zwaenepoel’s Blast protocol [4, 5], VMTP [6], and XTP [7]. Present-day “heavyweight” protocols such as OSI place notable performance con- straints on distributed applications. Svobodova [8] surveys the status of transport proto- cols (OSI and others) running on LANs and finds throughputs up to a maximum of about 2 Mbits/s (other references are found in her paper). The goal of the current lightweight 1