Author: Johannes Luethi

 Title: Analysis of Queueing Network Models for Computer and
	Communication Systems with Workload Uncertainties and Variabilities

PhD Thesis (in applied mathematics), University of Vienna, Austria.
 

Abstract:

Increasing complexity of computer and communication systems, especially
of distributed and parallel systems, necessitates effective tools for
understanding their behavior and for predicting their performance.
An important and popular formalism for the modeling of computer systems 
is the approach of queueing network models.  Conventional evaluation 
techniques for networks of queues, like e.g. the well-known Mean Value 
Analysis (MVA) algorithm, accept single values as input parameters (e.g., 
mean service times) and produce single mean values as performance 
measures (e.g., the mean system throughput).  However, the exact value of 
every parameter of the model may not always be known.  Furthermore, 
system parameters are often subject to variabilities.  Ignoring such 
uncertainties and variabilities in the workload may lead to inaccurate 
and incorrect analysis results.  Thus it is important to consider such 
effects in the workload characterization of the model.

In this work, queueing network models of systems which exhibit 
uncertainties and variabilities, are characterized by lists of vectors 
with parameter intervals.  A probability of occurrence is associated to 
every entry in such a list.  Sources for such parameter lists may for 
example be parameter histograms, clustering techniques, or approximations 
of parameter distributions.  Existing analysis techniques have to be 
adapted to handle this type of workload characterization.  In a first 
step, conventional techniques are adapted to interval parameters, which 
allows computing lists of intermediate interval results with associated 
probabilities of occurrence.  To obtain a better overview, these 
intermediate intervals are aggregated to representations which allow an 
easier interpretation of the results.  In this dissertation, the 
adaptation to the proposed workload characterization is presented for the 
MVA algorithm as well as for performance bounds, bottleneck-, and 
modification analysis techniques.