Year: 2007
Type: dissertation
Abstract: Large-scale distributed computing platforms have become widely used due to improvements in hardware and software.To evaluate algorithms for those platforms, one usually conducts experiments in simulations.While simulation has been employed in various areas such as microprocessor design and network protocol design, simulation techniques for large-scale distributed applications are still in their infancy.In this thesis, we focus on network simulation, which is one of the most important components of simulation for large-scale distributed applications.We first investigate the trade-offs between simulation cost and accuracy for two commonly used network simulation methods, packet-level and analytical simulations.We do this by comparing three existing packet-level network simulation tools, GTNetS, SSFNet, and ns-2, with the analytical simulation models implemented in SIMGRID, a toolkit for simulating Grid applications.We find that the following three factors affect the accuracy of SIMGRID when compared to packet-level simulation: (1) the size of the data transfered by TCP flows, (2) the complexity of the network topology, and (3) the network contention.Our results provide a new understanding of the relative merit of analytical network simulation versus packetlevel simulation.We then integrate GTNetS into SIMGRID, in order to enable the SIMGRID users to switch between packet-level and analytical network simulations and make their own decision regarding the costl accuracy trade-offs.The two main challenges are the translation between simulation objects in the two simulators, and the addition of suspendable/resumable simulation and of speculative simulation capabilities to GTNetS.We address the latter by modifying the GTNetS source code, and address the former by implementing an API on top of the modified GTNetS.The end result for SIMGRID users is a new ability to switch easily (i.e., with no user code modification) between analytical and packet-level simulations.
Author: Kayo Fujiwara
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2310074916