TOWARDS A HIGH SPEED MAN ARCHITECTURE zyxwv Ahmed Tantawy', Hanafy Meleis', Magda El Zarki2 and Gowthami Rajendran2 I IBM Research Division, T. J. Watson Research Center, Yorktown Heights, NY zyxw * University of Pennsylvania, Department of Electrical Engineering, Philadelphia, PA Abstract Many real applications could be implemented if very high speed communication systems were available. High bandwidth network infrastructures exist today. However, the end-to-end throughput of actual communication sys- tems is far less than what the physical links can handle. This paper attempts to identify the bottlenecks in current network architectures and presents the outline of a framework for the design of an architecture and protocol suite for high speed Metropolitan Area Networks (MANs). 1. Introduction It has been claimed that optical fibers have revolu- tionized the communications industry. And indeed to a great extent they have. The bit error rate has fallen from 10-6 in copper coaxial cables to in optical fibers. The transmission rates have increased from a few Mb/s in coaxial cables to the order of Gb/s in fiber optic links. Many applications exist which can utilize such high transmission rates. Examples of such applications in- clude medical imaging, advanced CAD/CAM, videophone and videomail, computer center consol- idation, etc. This abundance of bandwidth seems to im- ply that all these applications can now be realized. This however, is a fallacy, because while the transmission speeds over the physical links have increased dramat- ically, the end to end throughput visible to the applica- tions has only increased marginally zyxwvutsrqp [5,23]. So, how can the throughput be improved? Because of the complex structure of a communication network, this question is not simply answered. Current research in this direction can be broadly classified into two cate- gories [23]: zyxwvutsrqpo 1) Modification zyxwvutsrq of current protocol designs. Some design- ers suggest that current protocols are not suitable for high speed networks because they were developed for a different transmission environment, one in which the bit error rate was high and where the ratio of transmission rate to processing speed was in favor of the latter. Consequently, some schemes used for error control and flow control which performed efficiently in low speed networks fail miserably when applied directly to high speed networks. For instance, simple calculations show that using the basic HDLC protocol in very high speed networks would result in having the window closed for approximately 99% of the time. 2) Improved implementation of current protocols. The other school of thought advocates that efficient imple- mentation of current protocols can lead to a major in- crease in throughput. It has been shown that the implementation of various protocols has a major effect on throughput [23]. Most protocol implementations to- day use common general purpose microprocessors to carry on the computations needed. Unfortunately, the speed of the currently available processors are a serious bottleneck in high speed networking. Some important issues that affect the performance of networks are the hardware versus software implementation of the various functions and protocols. Some tests conducted along the lines of the tech- niques mentioned above did not produce very promising results [22]. Hence, we believe that serious consideration should instead be given to the restructuring of the entire communication systems architecture in order to make the best use of the current technology at the user level. In fact, one should view the communication system as part of an environment designed to carry out given distributed applications. These two schools of thought will be examined in this paper with particular focus on Metropolitan Area Net- works (MANs). These networks are regarded as the sol- ution for high speed communications within certain geographical limits without having to deal with long dis- tance carriers and their restrictions, regulations and sometimes conflicting interests. It must be noted that the bulk of the research in MANs has not focussed on end to end throughput, instead efforts have been directed towards improving the Media Access Control (MAC) 20.1.1. CH2655-9/8910000-0619 $1 .OO zyxwvuts 0 1989 IEEE zyxwv 061 9