High Performance Computing for Wireless Telecom Research Maziar Nekovee BT Research & University College London maziar.nekovee@bt.com T. Hewer, R. Saksena, P. V. Coveney (UCL) Keith Briggs, S. Kawade, M. Fitch (BT)
3G/4G Today s wireless telecom networks Control Server Cellular WiFi IP backhaul Network IEEE Gateway WiMax Sensors Public Internet WiFi Corporate Computers Application (Web) Servers Email YouTube Social networks. Ad-Hoc
Computational modelling for wireless telecom research Network planning and optimisation base station positioning, frequency planning, transmit power control (Green ICT) Technology evaluation/exploration: WiFi, WiMAX, 3G/4G, MIMO, Cognitive Radio S. Kawade et al, 2009 (submitted) New frontiers in mobile wireless communications and networking transport, health, wireless grids, mobile social nets Agent-based business modelling Spectrum auctions (bidding strategies), customer s choice dynamics (complex social nets)
Combinatorial optimisation Frequency planning Routing in mesh networks Spectrum auction modelling K. Briggs, IEEE VTC 2009 Soho (Central London), 3 WiFi channels 11 Mbps, 5.5 Mbps, 2Mbps, 1Mbps number of WiFi access points
Simulations of V2V communication networks for future intelligent transport car-to-car messaging WiFi-equipped car Source: S. Euobank Virginia Tech, 2006
Coupled simulations smallest time scale is second smallest time scale is nanosecond bottleneck
Congestion reduction using V2V messaging TraffiCom: Car-following traffic simulator + link-level wireless simulator Accident Nekovee and Bogason, IEEE VTC, 2007 Nekovee, Proc. IET on Intelligent Transport, 2008 Hewer and Nekovee, IEEE Workshop on V2V Comm., 2008
Hewer and Nekovee 2009, submitted
Inside a high-fidelity wireless discrete event simulator: NS2/NS3, OMNET++, Qualnet, Opnet, Application Pkt Application Application LL LL LL MAC802.11 Sender MAC802.11 Recv MAC802.11 Victim RF WirelessPhy RF WirelessPhy Pkt RF WirelessPhy Pkt Wireless Channel 10
Scalability of wireless network simulators Parameter searches/monte Carlo runs Scalability studies prior to large-scale deployment Coupled simulations (real-time or better is required) 10 minutes NS3 simulations: 10 minutes of V2V messaging in a stretch of highway
Scalability of wireless network simulators Fundamental bottleneck is sequential architecture parallel and distributed simulations on HPC is a must
Parallel parameter exploration/optimisation on HPC: Legion UCL Hewer, Nekovee, Saksena, Coveney, AHM 2008
Domain decomposition parallel simulations Coupling between wireless devices is very local (100-1000 m for WiFi) Divide the network into geographical sub-domains and assign each subdomain to a single core Each core simulates communication and networking within its own subdomain Inter-processor communications at boundaries via MPI Complex synchronisation schemes are needed to handle asynchronous dynamics of discrete event simulations CMP Mesh Link CAP CGW Local events only processed if they are not ahead of any other global event that may affect them. sub-domains Nekovee, Saksena, FGCS (2008)
Parallel performance of the simplified wireless code on HPC: Teragrid Saksena, Nekovee, Coveney, 2009 Simulation of wireless computer virus spreading in a 64000 node mobile WiFi network: Only MAC and PHY layer included (highly simplified)..
Conclusions and future work Computational modelling and simulation is becoming increasingly more important in wireless telecom research (both academia and industry). HPC is absolutely essential to meet computational challenges of complexity and scalability; work by us and others (e.g. Motorola Research, US) is demonstrating this. The class of computational problems that need to be tackled: parallel discrete event network simulations, combinatorial optimisation, agent-based modelling pose new challenges in design of parallel, distributed and numerical algorithms that we aim to address with the help of the HPC/NA community.