Processor Management for Adaptive Applications

Hans Domjan, Ph.D. Thesis

Despite the ever increasing speed of hardware, today's multimedia-rich
applications on a user's workstation and distributed computing in the
Internet (Web-Servers providing dynamic content like online time
tables) manage to consume the entire resources of networks or end
systems. While it is economically infeasible to size resources for
peak usage, end users nevertheless expect a predictable service from
the applications, i.e., the result should be produced within a certain
time limit. A solution to this problem is to design applications so
that they adapt to available resources: in a resource contention
situation, they lower the quality of their result and thus consume
fewer resources, but produce the result on time. If there are ample
resources, they are able to produce full-quality results, and still
meet the time limit.

This dissertation investigates end system support to provide
predictable service that matches the needs of adaptive applications
with regard to the resource CPU (Central Processing Unit). It presents
the {\em RV-Scheduling System}, which provides resource reservations
on top of a best-effort operating system, thus combining the
flexibility and versatility of such OSes with resource guarantees as
found in real-time systems.

Based on three example adaptive applications, we introduce a model of
adaptive applications and derive their requirements with regard to end
system CPU management. Subsequently we introduce the design (including
the application programming interface API) and implementation of the
proposed RV-Scheduling System, which fulfills these requirements. Two
key ideas of the implementation consist of an {\em objective function
for cumulative resource consumption} combined with a {\em reactive
scheduling mechanism}. A thorough evaluation of the scheduler shows
that the RV-Scheduling System provides---despite its integration into
a best-effort operating system---resource reservations in a quality
and with a predictability which is useful for adaptive applications,
yet has negligible additional overhead. Applications using the
RV-Scheduling System experience a considerable increase in
user-specific quality metrics compared to best-effort
scheduling. Furthermore, the RV-Scheduling System enables the
operating system to make more efficient use of the resource CPU, and
allows applications to gracefully and dynamically decrease their
service quality without a-priori knowledge of the available
resources. Thus the RV-Scheduling System is considered a valuable
addition to today's best-effort operating systems.

(C) Hans Domjan, 2001