Enhancing Realism And Scalability In Network Testbeds
PhD Dissertation, University of Utah. May 2010.
Network emulation has become an indispensable tool for the conduct of research in networking and distributed systems. It offers more realism than simulation and more control and repeatability than experimentation on a live network. However, emulation testbeds face a number of challenges, most prominently realism and scale.
Because emulation allows the creation of arbitrary networks exhibiting a wide range of conditions, there is no guarantee that emulated topologies reflect real networks; the burden of selecting parameters to create a realistic environment is on the experimenter. While there are a number of techniques for measuring the end-to-end properties of real networks, directly importing such properties into an emulation has been a challenge. Similarly, while there exist numerous models for creating realistic network topologies, the lack of addresses on these generated topologies has been a barrier to using them in emulators.
Once an experimenter obtains a suitable topology, that topology must be mapped onto the physical resources of the testbed so that it can be instantiated. A number of restrictions make this an interesting problem: testbeds typically have heterogeneous hardware, scarce resources which must be conserved, and bottlenecks that must not be overused. User requests for particular types of nodes or links must also be met. In light of these constraints, the network testbed mapping problem is NP-hard. Though the complexity of the problem increases rapidly with the size of the experimenter’s topology and the size of the physical network, the runtime of the mapper must not; long mapping times can hinder the usability of the testbed.
This dissertation makes three contributions towards improving realism and scale in emulation testbeds. First, it meets the need for realistic network conditions by creating Flexlab, a hybrid environment that couples an emulation testbed with a live-network testbed, inheriting strengths from each. Second, it attends to the need for realistic topologies by presenting a set of algorithms for automatically annotating generated topologies with realistic IP addresses. Third, it presents a mapper, assign, that is capable of assigning experimenters’ requested topologies to testbeds’ physical resources in a manner that scales well enough to handle large environments.