Performance Tradeoffs in Priority Scheduling
Jeffrey P. Buzen <email@example.com>
880 Winter St.
Waltham, MA 02454
Priority scheduling always involves tradeoffs - high priority customers do
better, while low priority customers do worse. The equations that
characterize these tradeoffs can be found in advanced texts on queuing
theory. However, because these equations are so complex, it is difficult to
gain an appreciation of the critical underlying issues simply by examining
the corresponding closed form analytic solutions.
This tutorial presents an intuitive, easy to understand explanation of the
basic factors that regulate performance in systems with priority scheduling.
The emphasis is on intuition and insight rather than mathematical details.
Nevertheless, the tutorial will provide enough information to enable
attendees to carry out "back of the envelope" computations and develop
simple programs for analyzing specific problems in priority scheduling.
The basis for this tutorial is provided by Kleinrock's well known
conservation laws for priority systems. Several variants of these laws will
be considered. In addition, the implications of achievability constraints
and the transparency principle will be examined from both an analytic and a
A series of numerical examples and sample applications will be used to
illustrate the main points. The first involves "class of service/quality of
service" scheduling in Web server farms. The second involves the goal
oriented dispatcher (Workload Manager) developed by IBM for their largest
mainframe systems. In addition to providing concrete examples of the
necessary computational procedures, the applications also incorporate
unexpected and counter-intuitive results.
Who should attend?
Anyone interested in learning more about the performance of priority based
systems. No prior knowledge of specific results from queuing theory will be
required, but basic familiarity with performance concepts and algebraic
computations will be assumed. This tutorial should be of interest to
practitioners responsible for the management of system performance,
educators seeking classroom material on this topic, and system designers
responsible for developing dispatching algorithms used in computer systems,
network routers and a variety of other applications.
Jeffrey P. Buzen has been a leader in the field of computer performance
analysis since the early 1970's. His career spans an unusually broad range
of activities including the creation of fundamental mathematical models and
analysis procedures (central server model, convolution algorithm,
operational analysis), the development of practical tools based on these
theoretical advances (BEST/1, CAPTURE/MVS), and the founding of a
commercially successful software company (BGS Systems) that extended,
supported and marketed these tools for more than two decades. Throughout
his career, he has published extensively, lectured widely and held
leadership positions in several professional societies including ACM
Sigmetrics, IFIP Working Group 7.3, and CMG.
Jeff Buzen is currently President of the Computer Measurement Group
(www.cmg.org), a professional society for computer performance
practitioners. He also holds a part time appointment as a Consulting
Scientist at BMC Software. His Ph.D. in Applied Mathematics is from Harvard
University (1971). From 1971 to 1976 he held concurrent appointments as a
Lecturer in Computer Science at Harvard and a Systems Engineer at Honeywell.
Among his students at Harvard were some of the most influential figures in
the computer industry: Robert M. Metcalfe (Ph.D. 1973), inventor of
ethernet, John M. McQuillan (Ph.D. 1974), developer of dynamic routing
algorithms used in the ARPAnet/Internet, and William H. Gates, co-founder of
In 1975, Jeff Buzen co-founded BGS Systems along with his classmates Robert
P. Goldberg and Harold S. Schwenk. He served as Chief Scientist and Senior
Vice President at BGS until its merger with BMC Software in 1998.