Affiliations: Advanced Systems Division, National Institute of Standards and Technology, Building 225, Room A231, Gaithersburg, MD 20899, USA, Tel: 301 975 2904, Fax: 301 840 1357, [email protected]
Note: [] This work was partially sponsored by the Advanced Research Projects Agency.
Note: [] This National Institute of Standards and Technology contribution is not subject to copyright in the United States. Certain commercial equipment, instruments, or materials may be identified in this paper to adequately specify experimental procedures or to provide examples. Such identification does not imply recommendation or endorsement by the National Institute of Standards and Technology, nor does it imply that materials or equipment identified are necessarily the best available for the purpose.
Abstract: The utilization of the inherent bandwidth in high-speed networks is often obstructed by the implementation of the protocols. NIST and ARPA developed VLSI measurement components that permit researchers and product developers to better understand computing and communication bottlenecks by accurate measurement techniques. To illustrate and promote the use of such public domain components, this paper uses image transfer as an example application via a commercial UNIX (UNIX is a trademark of X/Open) (BSD 4.3.1 variant) implementation of the TCP/IP protocol over an Ethernet (10 Mbits/s), and a HIPPI (800 Mbits/s) medium. The only modification done to this commercially available kernel was to add the simple probe code necessary to interface to the NIST MultiKron (MultiKron is a registered trademark of NIST) measurement chip. The accurate data obtained for the various aspects of this protocol implementation clearly illustrate the major bottlenecks, provide insight with regard to scaling, and determine upper bounds on performance. This paper does not, however, attempt to draw any conclusions with regard to the abstract TCP/IP protocol itself or the generic services of the UNIX operating system, rather it provides a system-wide perspective on how a wide range of design options and implementation parameters impact the effective throughput to the application.
