Summary: Measured capcacity of Ethernet

This primary purpose of this paper is to dispel some theoretical myths
about Ethernet and confusions about Ethernet capacity under heavy loads by
measuring Ethernet performance in actual implementations.

The paper is roughly divided into four sections. Firstly, a brief summary
of theoretical studies relevant to Ethernet is presented. It also
summarizes various parameters affecting the Ethernet performance.
Secondly, it explains why theoretical studies fail to study the real
performance of Ethernet. The reasons why they failed vary from unrealistic
assumptions, and operating in the wrong regime, to using approximations
for offered load.

Thirdly, the paper defines the measurement environment and methodology for
studying the Ethernet performance. The performance indicators such as
response time, average packet delay, fairness and efficiency are measured
by varying packet arrival distribution, the packet sizes, number of hosts
(offered load) and size of Ethernet segments(longnet/shortnet etc). One
important observation is that these simulations are performed in an
environment which Ethernet is designed for. For e.g. To approximate real
traffic, the authors used many minimum size packets, some maximum length
packets and a few intermediate size packets as opposed to using only fixed
size packets. The results prove that Ethernet is indeed capable of good
performance for typical applications, even at high loads. Finally, it
summarizes the lessons learned from theoretical studies. It also briefly
outlines common problems associated with the implementations such as
incorrect hardware design and jabbing transmitter etc. and suggests ways
to avoid future problems.

The conclusions about the Ethernet performance are based on the studies
performed in an environment that has 24 nodes and four Ethernet segments.
In real world scenario, the no of hosts are usually more than 24(in few
hundreds). The authors could have increased the no of hosts to a
reasonable number close to typical Ethernets to understand scalability of
the results. One important question paper does not answer is that: if the
no of hosts (N) is high, how does Ethernet work if there are lots of
minimum size packets? Also, the paper does not the clearly explain the
relationship between offered load and number of hosts (I am guessing that
it could cause some deviation in their results).
Received on Tue Sep 19 2006 - 01:30:24 EDT