Improved performance for network simulation

Over the course of designing and implementing two discrete event simulators, the commercial simulator packages CSIM and DesmoJ were leveraged to allow for rapid development of both wired and wireless network models. However, the two resulting simulators demonstrated poor scalability due to the use of multi-threading to maintain state for simulation elements. By using a simple single-process discrete event simulation engine, the running-time showed a marked decrease when compared to multi-threaded simulators. In one case study, we simulate a simple two-link MPLS network which employs two congestion control mechanisms for inelastic traffic, namely preemption and adaptation. Performance metrics measured include: the per-class blocking probability, customer average fraction of time streams travel on the preferred path, customer average fraction of time at the maximum subscription rate, the customer average rate of adaptation, and the time average rate of preemption. We compare the performance of preemption and adaptation individually and collectively against the base case where neither congestion mechanism is used. At the cost of increased number of rate adaptations and preemption events for a range of regimes, we show that the combined use of preemption and adaptation improves the quality of service and alignment of high priority traffic while increasing the effective network capacity. As a performance enhancement to the simulator developed to conducted these experiments, we switched to a single-process discrete evnt simulation engine in place of multi-threaded simulator. We note a large improvement for the running time as the simulation time and capacity increase. A second case study was conducted on a wireless simulator. In an effort to simplify the simulator and improve performance we again moved from a commercial thread-based simulator (CSIM) to a single-process discrete event simulation engine. Results of the runningtime vs network size for the single-process simulator showed a constant-time improvement over the thread-based simulator. To further improve performance, a complementary technique known as model abstraction is also applied. Model abstraction is a technique that reduces execution time by removing unnecessary simulation detail. In this thesis we propose three abstractions of the IEEE 802.11 protocol. The Goodput Ratio vs Transmission Power and End-to-end delay vs offered load performance metrics are compared against the OPNET commercial simulator.