Review: Measured Capacity of an Ethernet

Paper: Measured Capacity of An Ethernet: Myths and Reality

Name: Robert Danek.
Course: CS2209, Fall '06.

    This paper has three main goals. First, it wants to dispel a number
of myths that have arisen about the performance of Ethernets from
theoretical studies. Second, the authors conduct a number of experiments
to demonstrate the performance of a real Ethernet. Finally, the paper
provides some guidelines for the implementation of Ethernets to ensure
maximum possible performance.

    One myth arising from the misinterpretation of theoretical studies
is that Ethernets "are saturated at offered loads of 37%." The
unrealistic assumptions that certain theoretical studies make include:
assuming hosts are at identical and maximum distances away from each
other; assuming packet lengths are unimodal; and assuming no buffering
occurs. Other problems with theoretical studies are that some of them
study non-persistent or p-persistent CSMA with or without CD, while
standard Ethernet is 1-persistent CSMA/CD.

    The experiments that the authors conduct involve heavily loading the
Ethernet that they test. They want to study the behaviour exhibited by
the network when intermittent overloads occur. To do this, they apply a
steady overload. They find that the network's performance is highly
dependent on the packet length and the time to detect collision.

    The authors conclude by providing insights on how a real Ethernet
should be used. These insights are obtained from both theoretical and
experimental studies. They also point out that network performance can
be hampered by other bottlenecks besides the network itself, including
processor and software limitations.

    Overall I thought this was a good paper. It emphasized the point
that theoretical analyses of real systems, while they can provide some
level of insight about the system, cannot replace empirical studies.
Received on Mon Sep 18 2006 - 15:30:03 EDT