1 Queuing Theory vs. System I dentification Network Configurations - PDF document

Research Background Research Background Measurement- Measurement -Based Modeling of Based Modeling of receiver sender I nternet Round- I nternet Round -Trip Time Dynamics Trip Time Dynamics If characteristic of packet transmission delay packet transmission delay is is If characteristic of using System I dentification using System I dentification known… known … Can improve QoS QoS (Quality of Services) (Quality of Services) Can improve Hiroyuki Ohsaki Ohsaki Can realize efficient congestion control Hiroyuki Can realize efficient congestion control Department of Information Science and Dynamics of packet transmission delay Dynamics of packet transmission delay Department of Information Science and Technology, Osaka University, Japan Technology, Osaka University, Japan How packet transmission delay is affected by packet How packet transmission delay is affected by packet transmission rate transmission rate oosaki@ist.osaka- oosaki@ist.osaka -u.ac.jp u.ac.jp http://www.anarg http://www. anarg. .jp jp/ / Research Objectives Black- -Box Modeling Box Modeling Research Objectives Black Model dynamics of packet transmission delay Model dynamics of packet transmission delay View network as a View network as a black box black box sender black box Measure input and output to network Measure input and output to network input receiver Estimate model parameters by system identification Estimate model parameters by system identification output Validate model accuracy Validate model accuracy Using two validation methods Using two validation methods Definition of input and output Definition of input and output Comparison with simulation (time domain) Comparison with simulation (time domain) Comparison with spectral analysis result (frequency domain) Comparison with spectral analysis result (frequency domain) input packet transmission rate round-trip time output ARX （ （ eXogenous ） ） Auto- -Regressive Regressive eXogenous Model ARX Auto Model Modeling Procedure Modeling Procedure noise e(k) (effect of other traffic, etc) input data model ARX model u(k) y(k) output data parameters ARX model system identification input output (transmission rate) (round-trip time) 1. Measure input and output data = - - - - - y ( k ) a y ( k 1 ) ..... a y ( k n ) 2. Determine ARX model parameters from 1 n a a measured input and output data + - + + - + b u ( k 1 ) ..... b u ( k n ) e ( k ) 1 n b 3. Validate accuracy of ARX model obtained b Determined by system identification 1

Queuing Theory vs. System I dentification Network Configurations Queuing Theory vs. System I dentification Network Configurations Build model in three network configurations Build model in three network configurations queuing theory queuing theory system identification system identification Complexity of network topology Complexity of network topology Effect of noise noise (e.g., background traffic) (e.g., background traffic) Effect of modeling approach modeling approach white box white box black box black box Location of bottleneck link Location of bottleneck link dynamics cannot model can model dynamics cannot model can model N1: LAN N1 : LAN Simple network topology Simple network topology control theory control theory no no yes yes N2: WAN N2 : WAN applicability applicability Complex network topology Complex network topology statistical statistical can model can model cannot model cannot model Access link is bottleneck Access link is bottleneck characteristic characteristic N3: WAN N3 : WAN model information required not required model information required not required Complex network topology Complex network topology Access link is not bottleneck Access link is not bottleneck model accuracy model accuracy good good ? ? Network N1 Network N2 Network N1 Network N2 bottleneck Internet sender receiver sender ISP 100Mbps access link receiver FTP server FTP client Mbps ms Mbps ms (packet transmission rate) 0.08 500 (packet transmission rate) 120 0.44 (Round-Trip Time) (Round-Trip Time) 0.06 0.43 80 Output Input Output Input 0.04 0.42 40 0.02 0 0 150 0 0.41 2000 2150 2000 2150 2000 2150 2000 2150 slot slot slot slot sampling interval = 125ms sampling interval = 0.9ms Validation using Simulation Validation using Simulation Network N3 Network N3 (Comparison in Time Domain) (Comparison in Time Domain) Internet bottleneck? sender input data 1 output data 1 receiver (packet transmission rate) Mbps ms 20 80 ARX model (Round-Trip Time) 15 60 Output input data 2 model output Input 10 40 Compare output data (not used for model 5 creation) with simulation output 0 0 20 2000 2150 2000 2150 slot slot output data 2 sampling interval = 6ms 2

Validation using Board Diagram Validation using Board Diagram (Comparison in Frequency Domain) Network N1: Simulation (Comparison in Frequency Domain) Network N1: Simulation sender receiver unknown 100Mbps ω ω + ϕ sin( t ) A sin( t ) FTP server FTP client ϕ A gain phase measured output model output Round-Trip Time obtain frequency characteristic 0.42 ARX model from ARX model estimate frequency characteristic input data 0.415 using spectral analysis output data Compare frequency characteristics of ARX model 0.41 and one estimated using spectral analysis 2250 2270 2280 2290 2300 2260 slot Network N1: Board Diagram Network N2: Simulation Network N1: Board Diagram Network N2: Simulation -2 10 bottleneck ARX model Internet amplitude spectral analysis sender -4 10 receiver -6 450 10 -2 -1 0 1 measured output 10 10 10 10 model output 400 Round-Trip Time 0 350 phase (degree) 300 100 250 200 200 -2 -1 0 1 10 10 10 10 150 2250 2260 2270 2280 2290 2300 slot frequency (rad/s) Network N2: Board Diagram Network N2: Board Diagram Network N3: Simulation Network N3: Simulation 4 10 bottleneck? Internet amplitude sender 3 10 receiver ARX model spectral analysis 40 2 10 -2 -1 0 1 10 10 10 10 Round-Trip Time 35 0 phase (degree) 30 200 25 measured output 400 model output -2 -1 0 1 10 10 10 10 20 2250 2260 2270 2280 2290 2300 slot frequency (rad/s) 3

Network N3: Board Diagram Conclusion Network N3: Board Diagram Conclusion 2 10 ARX model Model dynamics of packet transmission delay Model dynamics of packet transmission delay spectral analysis amplitude 0 View network as a black box View network as a black box 10 Measure input and output data Measure input and output data Determine model parameters by system identification Determine model parameters by system identification 2 10 -2 -1 0 1 10 10 10 10 Validate model accuracy Validate model accuracy 0 Simulation (comparison in time domain) Simulation (comparison in time domain) phase (degree) Bode diagram (comparison in frequency domain) Bode diagram (comparison in frequency domain) 500 Show effectiveness of black- Show effectiveness of black -box modeling box modeling When network is not so noisy not so noisy When network is 1000 -2 -1 0 1 10 10 10 10 frequency (rad/s) Future Works Future Works More accurate modeling of packet transmission delay More accurate modeling of packet transmission delay Model structures Model structures Parametric model Parametric model Non- -parametric model parametric model Non Linearity Linearity Linear model Linear model Non Non- -linear model linear model Noise assumption Noise assumption white noise white noise colored noise colored noise Design a delay Design a delay- -based congestion control using control based congestion control using control theory theory More info: http://www.anarg anarg. .jp jp/ / More info: http://www. 4