Scott Weber
CSE 498

Review of "Measured Capacity of an Ethernet: Myths and Reality"

This paper examines the performance of a standard Ethernet and the properties that of the network that effect the performance.  The authors also spend effort to dispel myths about Ethernet performance that have become common due to unclear and confusing previous research.  Many of his complaints about previous research is that it is largely based in theory and tend to examine how an Ethernet would behave in an unrealistic situation; for example, using a constant packet length or non-standard implementation of the protocol.

The performance testing was done on a real Ethernet, using several identical machines connected to the Ether.  Several conclusions are made from the results obtained.  A longer average packet length yields a higher usage in bits per second, but a lower usage in packets per second.  More hosts on a network tends to increase the fairness of usage, but increase the transmission delay.  Mixed amounts of short and long packets provides better utilization than just short packets.  The Ethernet is only as good as the weakest implementation on the network.

In section 2.1, when describing the parameters that are user determined, the authors state that "no single choice of these parameters can be viewed as typical" and then states just two paragraphs later that "typical installations have..."  There are two problems with this.  The first, and most obvious is that the second statement is in direct conflict with the first.  The second is that the source of this data is not stated.  What was done to obtain these typical values?  Were network administrators surveyed, implementations of the protocol examined, or wild guesses made?

When testing the performance of the Ethernet, the authors chose to run three tests outside the standard.  The length of the packets used was much longer than the maximum length as defined in the standard.  Earlier in the paper, the authors blame the confusion related to previous research on this very same issue: testing outside the standard.  No reason is given for testing outside the standard, and it seems these results can be just as misleading as the results of previous research.

Later in the performance analysis, when examining Figure 3-6, the authors state that "the high variance for 64-byte packets may be an experimental artifact."  No reason or inclination is given to back up this statement.  What made the authors believe it may be an artifact and not valid result?

Overall, this is a good paper.  It is useful for showing the capabilities and limits of an Ethernet and gives some hints on how one might tune a network to get optimal performance.

REFERENCE
David R. Boggs, Jeffrey C. Mogul, and Christopher A. Kent. "Measured Capacity of an Ethernet: Myths and Reality." Proc. of ACM SIGCOMM, August 1988, and Computer Communication Review 18(4).