Network Computer System Performance Evaluation

Daniel J.-C. Folly

Semester Thesis Winter 1998-99
Supervisors: C. Kurmann, Prof. T. Stricker
Institute for Computer Systems, ETH Zürich


Fast communication will be one of the key technologies in the uprising century. Development of new multimedia applications needs a faster access to a huge amount of data. Videoconferencing and interactive television could be technologies that will change our life. Besides new multimedia technologies online-distribution of software is becoming more essential. For such huge data the existing network technologies doesn't match any more.

I would like to make a comparison of the speed of a network-link between a 10MBit, 100MBit and a 1000MBit Ethernet network.


One of the key results is the distribution of the packet sizes. In a measurement in a real network I could discover that 34% of the observed packets were only 64 Bytes big. On the other side one half of the packets were between 1024 and 1518 Bytes. That means that there exists either very little or very big packets in a real office network.

Another key result in consideration of this bimodal distribution of packets sizes is the fact, that faster network reach there maximal throughput in a much slower way. In a 10Mbit Ethernet the max. throughput is approximated in an exponential manner. In the tested 1000Mbit Ethernet the speed for small packets growth almost linear.

The throughput diagram of the 10Mbit link is very typical. After the observed throughput reaches the maximum the curve remains almost constant with some periodical minor notches at the top level.

The same diagram for the 100Mbit link is much more chaotic. Blocks bigger than 4381 Bytes results in either a good or a bad throughput rate. After a known limit the throughput remains at a low level. This indicates a problem in either the used software or in one of the hardware parts.

The throughput of the 1000Mbit is different. After an increase to a top level the throughput collapses around the maximum TCP buffer size. For very big data blocks the throughput reaches again a high top level

The author can be reached at

[ CS-Department | Up ]
ETH Zürich: Department of Computer Science
Comments to Jacques Supcik <>
Date: 8. March 1999