Shashikiran B. Guruprasad
School of Computing
University of Utah, Salt Lake City, UT 84112
Discrete-event network simulation is widely used to rapidly design, evaluate, and validate new networking ideas as well as study the behavior of existing ones. It is characterized by complete control, absolute repeatability and ease of use while often sacrificing detail and realism to increase execution efficiency and the scale of models. Network emulation allows the study of applications running on real hosts and “somewhat real” networks. A key difference between the two approaches is that in the former, the notion of time is virtual and is independent of real time, whereas the latter must execute in real time. Typically, emulated resources are also distributed in nature. Thus, emulation gains realism while naturally foregoing complete repeatability; historically, emulation was also tedious to control and manage.
Integrated Experiments, where we spatially combine real elements with simulated elements to model different portions of a network topology in the same experimental run, enable new validation techniques and larger experiments than obtainable by using real elements alone.
In this thesis, we present a system in which we employ multiple loosely coordinated simulator instances running on distributed PCs in real-time to model the simulated portion of a network topology. Our key design techniques are to perform optimistic automated resource allocation, and to use feedback to adaptively allocate simulated resources in order for the simulators to run in real-time. Multiple simulator configurations specific to a resource assignment are automatically generated from an experimenter configuration which is agnostic to the details of the physical realization. The entire system is highly automated and is available for production use in Emulab.
The full thesis is available in these formats:
|Eric Eide <email@example.com>||Last modified: Sat Mar 14 11:11:00 MDT 2009|