While considerable research has gone into investigating various
networking and operating system mechanisms for supporting the
transfer and playout of stored continuous media (e.g., audio and video),
very little information is available about how users actually
{\it use\/} such systems.
Understanding how users interact with such a system --
i.e., developing a user workload characterization --
is crucial in designing and evaluating efficient
continuous media (CM) resource allocation and access mechanisms.
We have designed and built an interactive
WWW-based, multimedia, client/server application,
known as MANIC (Multimedia Asynchronous Networked Individualized
Courseware), that streams synchronized CM (currently audio)
and HTML documents to remote users.
MANIC was used by more than 200 users
during the Spring  1997 semester to listen to, and view, the stored
audio lectures and lecture notes for a full-semester senior-level
course at the University of Massachusetts.
In this paper we provide empirical and analytic
characterizations of observed user behavior in  MANIC. We characterize
both session-level behavior (e.g., the length of individual sessions)
as well as {\it interactive\/} user behavior (e.g.,
the time between starting/stopping/pausing the audio within a session;
and the ``distance'' of rewinding, fast forwarding and indexed jumping
within the audio stream).
We also examine how well our measured data can be fit analytically by various
distribution functions. Finally, we
consider  the possible implications of our results for
CM system architecture.