Hybrid network simulation

This thesis is motivated by the need to speed up traditional simulation of complex queueing networks, which large (discrete) transaction populations can easily render infeasible. Replacing discrete flows by continuous (fluid) flows, specified in terms of flow rates, is an alternative simulation paradigm for realizing speedups. The two simulation paradigms (discrete and continuous) offer a fundamental complexity tradeoff: a large number of computationally "cheap" packet-events versus a much smaller number of computationally "expensive" fluid-events.; This thesis develops a hybrid network simulation (HS) paradigm , where discrete transactions (e.g., packet flows) and continuous transactions (fluid flows) co-exist seamlessly in hybrid networks within the discrete-event simulation paradigm. It then describes the implementation of the HS paradigm in telecommunications setting as a software environment, dubbed HNS (Hybrid Network Simulator), where the modeler can exploit the tradeoff between the statistical accuracy of packet flows and the speedup afforded by fluid flows to make a judicious selection of flow types. Several telecommunications transport protocols are implemented in HNS as compatible network-model versions (pure-packet, pure-fluid and hybrid). Hybrid simulation, however, introduces constraints (e.g., restriction to feed-forward topology), implementation complexities, and run-time problems. The thesis develops data structures and algorithms to mitigate these problems and enhance run-time performance.; Using HNS, extensive experimentation was conducted to study the accuracy and computational complexity of compatible network-model versions as function of network complexity and workload/bandwidth. The results support the contention that fluid and hybrid simulations are efficacious in that they closely approximate the statistics of compatible pure-packet simulation, but at a lower computational cost. Pure-fluid versions were found to deliver the largest (several orders of magnitude) speedups, whereas hybrid versions attain only modest speedups.; HNS was further used to study the accuracy and stabilization of IPA (infinitesimal perturbation analysis) gradients of performance metrics with respect to parameters of interest in the aforementioned network-model versions. The thesis extends fluid-based IPA derivatives to pure-packet flows and hybrid flows, and uses HNS to support the contention that compatible network-model versions yield close stable IPA gradients. These results hold out the promise of using IPA gradients in control applications and simulation-based optimization of packet networks.