A large-scale real-time network simulation study using PRIME

Jason Liu, Yue Li, Ying He
2009 Proceedings of the 2009 Winter Simulation Conference (WSC)  
We examine the capabilities of conducting network experiments involving a large-scale peer-to-peer web-content distribution network. Our study uses a real-time network simulator, called PRIME, running on EmuLab, which is a shared cluster computing environment designed specifically for network emulation studies. Our study is one of the largest network experiments that involve a real implementation of a peer-to-peer content distribution system under HTTP traffic from a public-domain empirical
more » ... load trace and using a realistic large network model. Our experiments demonstrate the potentials of real-time simulation for studying complex behaviors of distributed applications under large-scale network conditions. INTRODUCTION 1.1 Parallel Real-Time Simulation PRIME (2009), which stands for Parallel Real-time Immersive network Modeling Environment, is a high-fidelity parallel network simulator capable of running large-scale network models. The implementation of PRIME inherits most of our previous efforts in the development of Scalable Simulation Framework (SSF), a process-oriented and conservatively synchronized parallel simulation engine designed for multi-protocol communication networks. SSF can run on most platforms, including shared-memory multiprocessors and clusters of distributed-memory machines. The SSF simulation engine is ultra fast ) and has been demonstrated to be capable of handling large network models, including simulations of the Internet (Cowie et al. 1999, cellular systems (Delve and Smith 2001), wireless ad hoc networks (Liu et al. 2005 , Newport et al. 2007 , and wireless sensor networks . In order to support large-scale simulation, PRIME relies on advanced parallel simulation techniques to harness the collective computing resources of parallel computers for an increased event-processing power. For example, to achieve good performance on distributed-memory machines, PRIME adopts a hierarchical synchronization scheme to address the discrepancy in the communication cost between distributed-memory and shared-memory platforms (Liu and Nicol 2001). Further, PRIME implements the composite synchronization algorithm (Nicol and Liu 2002), which combines the traditional synchronous and asynchronous conservative parallel simulation algorithms. When used alone, the traditional approaches have been shown to be inefficient to deal with particular aspects of the model topology. The composite algorithm is able to efficiently simulate diverse network scenarios, including those that exhibit large variability in link types (particularly with the existence of low-latency connections), and in node types (especially for those with a large degree of connectivity). PRIME extends SSF with emulation capabilities, where unmodified implementations of real applications can interact with the network simulator that operates in real time. That is, the simulation time in PRIME can be made to advance synchronously with the wall-clock time. Traffic originated from the real applications is captured by PRIME's emulation facilities and forwarded to the simulator; the real network packets are treated as simulation events as they are "carried" on the virtual network and experience appropriate packet delays and packet losses according to the run-time state of the simulated network. We call such simulation-based emulation approach real-time simulation (Liu 2008). Since the virtual network can interact seamlessly with the real network appliances, PRIME appears to be indistinguishable from a physical network in terms of conducting real traffic. Liu, Li, and He Large-scale real-time network simulation requires simulation be able to characterize the behavior of a network, potentially with millions of network entities and with realistic traffic load-all in real time. On the one hand, we apply parallel and distributed discrete-event simulation techniques to speed up simulation of large-scale networks. On the other hand, we use models at different levels of abstraction to reduce the computational demand while maintaining a desired degree of accuracy (Liu 2006 . For example, our previous studies (Liu and Li 2009) have shown that the parallel hybrid traffic model implemented in PRIME, which combines a fluid-based analytical model based on ordinary differential equations with the traditional packet-oriented discrete-event simulation, can achieve a speedup of more than three orders of magnitude over packet-oriented simulation without significant loss of accuracy. PRIME's multi-resolution modeling capability allows network emulation at unprecedented scales. Through extensive studies, we have been able to successfully emulate many applications, including routing algorithms, transport protocols, content distribution services, web services, multimedia streaming, and peer-to-peer networks (e.g., , Erazo et al. 2009 ).
doi:10.1109/wsc.2009.5429678 dblp:conf/wsc/LiuLH09 fatcat:6kyuki2mdjcsle3klg6gqb4mxm