Measured Capacity of an Ethernet: Myths and Reality

Motivation: The paper attempts to make a realistic analysis of the
measured capacity of an Ethernet by performing measurements of network
performance and gives an overview of theoritical work that has attempted
to quantify the same.

Key Points:
1) Most of the earlier work (mainly theoritical) makes unreasonable
simplifying assumptions about the nature of the network (like constant
high load,constant packet lengths and non-standard implementation)
which renders their work not always applicable to Ethernet(unslotted,
1-persistent, CSMA method with CD and binary exponential back-off )
2) Theoretical analysis if needs to be applicable needs to take into
account the various variable parameters which makes it a difficult task
and atleast should be complimented by actual measurement studies for
verification of claims. This paper studies Ethernet under varying
combinations of packet lengths, network lengths and number of hosts.
3) The paper gives useful and practical advice on the use of ethernet
effectively which is well supported by measurements performed.
4) Failure of ethernet is mostly due to hardware or accidental high
loads contrary to the then existing belief of high loads. Most commonly,
software and design bugs result in such failures.
5) Ethernet turns out to perform pretty well even for high bandwidth
applications although performance degrades in cases of very large number
of simultaneous hosts and low latency requirement.
6) The paper gives attention to detail clearly demarcating the
difference between the usage of G and Q as offered load.
7) The applicability of these results needs to be revisited with higher
and higher bandwidth networks being supported as well as modifications
being done in order to make such high speed networks a reality.

Lesson learnt: Measurement is essential to backup theoretical claims
Received on Tue Sep 19 2006 - 02:29:08 EDT