Computer Laboratory

Technical reports

Measurement-based management of network resources

Andrew William Moore

April 2002, 273 pages

This technical report is based on a dissertation submitted by the author for the degree of Doctor of Philosophy to the University of Cambridge.

Abstract

Measurement-Based Estimators are able to characterise data flows, enabling improvements to existing management techniques and access to previously impossible management techniques. It is the thesis of this dissertation that in addition to making practical adaptive management schemes, measurement-based estimators can be practical within current limitations of resource.

Examples of network management include the characterisation of current utilisation for explicit admission control and the configuration of a scheduler to divide link-capacity among competing traffic classes. Without measurements, these management techniques have relied upon the accurate characterisation of traffic – without accurate traffic characterisation, network resources may be under or over utilised.

Embracing Measurement-Based Estimation in admission control, Measurement-Based Admission Control (MBAC) algorithms have allowed characterisation of new traffic flows while adapting to changing flow requirements. However, there have been many MBAC algorithms proposed, often with no clear differentiation between them. This has motivated the need for a realistic, implementation-based comparison in order to identify an ideal MBAC algorithm.

This dissertation reports on an implementation-based comparison of MBAC algorithms conducted using a purpose built test environment. The use of an implementation-based comparison has allowed the MBAC algorithms to be tested under realistic conditions of traffic load and realistic limitations on memory, computational resources and measurements. Alongside this comparison is a decomposition of a group of MBAC algorithms, illustrating the relationship among MBAC algorithm components, as well as highlighting common elements among different MBAC algorithms.

The MBAC algorithm comparison reveals that, while no single algorithm is ideal, the specific resource demands, such as computation overheads, can dramatically impact on the MBAC algorithm’s performance. Further, due to the multiple timescales present in both traffic and management, the estimator of a robust MBAC algorithm must base its estimate on measurements made over a wide range of timescales. Finally, a reliable estimator must account for the error resulting from random properties of measurements.

Further identifying that the estimator components used in MBAC algorithms need not be tied to the admission control problem, one of the estimators (originally constructed as part of an MBAC algorithm) is used to continuously characterise resource requirements for a number of classes of traffic. Continuous characterisation of traffic, whether requiring similar or orthogonal resources, leads to the construction and demonstration of a network switch that is able to provide differentiated service while being adaptive to the demands of each traffic class. The dynamic allocation of resources is an approach unique to a measurement-based technique that would not be possible if resources were based upon static declarations of requirement.

Full text

PDF (1.7 MB)

BibTeX record

@TechReport{UCAM-CL-TR-528,
  author =	 {Moore, Andrew William},
  title = 	 {{Measurement-based management of network resources}},
  year = 	 2002,
  month = 	 apr,
  url = 	 {http://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-528.pdf},
  institution =  {University of Cambridge, Computer Laboratory},
  number = 	 {UCAM-CL-TR-528}
}