Review: Measured Capacity of an Ethernet: Myths and Reality

 The paper attempts at clearing up several misunderstandings about Ethernet
capacity that have resulted from several theoretical studies. The authors
first clarify some of the misunderstandings of several theoretical studies
from 1976 to 1986. They also highlight important findings and results of the
analyses. The paper categorizes the reasons that lead to misunderstandings
into: different protocols, unrealistic assumptions, measuring the wrong
dependent variable, using the wrong independent variable and operating in
the wrong regime.
 Furthermore, in order to show that Ethernet capable of good performance,
the paper presents measurements of the behaviour of Ethernet while varying
packet lengths, network lengths and number of hosts. They find that results
are similar for fixed length packets on a long net and fixed length packets
on a short net, with the only difference being that shorter networks are
more efficient due to the reduced collision resolution time. They found
that, in general, bit rates increase with increasing packet size and
decrease with increasing number of hosts, fairness increases with the number
of hosts, packet rates peak at two hosts because there are no collisions
(the hosts synchronize quickly), the average delay increases linearly with
the number of hosts and finally the efficiency decreases with an increase in
the number of hosts. Finally, they also experimented with a more realistic
load consisting of varying ratios of short and long packets, and discovered
that many small packets and very few large ones resulted in better
utilization than small packets alone.
 The paper concludes by providing tips on good Ethernet implementation based
on theoretical analyses and performance measurements. For example, use short
cables, keep hosts per cable small, correctly implement the protocol, and
match transmitter and receiver performance.
 The authors also point out that throughput and bandwidth depend not only on
the network but also on memory system bandwidth and processor speeds. Also,
poor performance can result from poor implementations of Ethernet which can
lead to hardware failures (such as open or shorted cables) and accidental
high loads due to software bugs (such as incorrect back off algorithms).
 I think the paper is well-structured starting off off by defining several
parameters that affect performance (including both fixed parameters such as
bit rate, maximum propagation delay and persistence as well as user-defined
parameters such as packet length distribution, number of hosts and length of
cable) followed by a list of assumptions made during theoretical analysis to
simplify the model, for example, Poisson distribution of packet arrivals,
fixed number of buffers at the host and a balanced star topology. The paper
also defines the metrics used to measure performance such as average delay,
throughput, channel capacity, fairness and stability. All this provides the
background necessary to understand the rest of the paper.
Received on Mon Sep 18 2006 - 23:13:02 EDT