180 likes | 319 Views
Fluid-based Analytical Model in GloMoSim. Jeffrey Twu (jctwu@cs.ucla.edu) Tak Kin Yung (tkyung@cs.ucla.edu) March 21, 2001. Outline. Motivation Project Plan Challenges Implementation Details Work Progress Simulation Topology QualNet results Future Directions. Motivation.
E N D
Fluid-based Analytical Model in GloMoSim Jeffrey Twu (jctwu@cs.ucla.edu) Tak Kin Yung (tkyung@cs.ucla.edu) March 21, 2001
Outline • Motivation • Project Plan • Challenges • Implementation Details • Work Progress • Simulation Topology • QualNet results • Future Directions
Motivation • Misra et al.’s SIGCOMM ’00 paper • Fluid-based Analysis of a Network of AQM Routers Supporting TCP Flows with an Application of RED • Analytical model using SDEs • Jump process driven Stochastic Differential Equation (SDE) model of TCP flows with RED • SDEs transformed to ordinary differential equations • Ordinary differential equations can be easily solved numerically, using MATLAB for example
Motivation • Promise of modeling large number of flows without a significant increase in computational complexity • Key result is striking similarity of graphs of queue length vs. time • Scenario parameters (connectivity, link bandwidth and delays, direction of flows, RED parameters minth, maxth, etc.) configured identically in both the ns-2 model and analytical model
Flow set 4 Flow set 1 RED router 1 Flow set 2 Flow set 3 Flow set 5 Motivation • 5 sets of flows (40 flows per set)2 RED routers • Flow set 2 passes through both routers RED router 2
Project Goals • GloMoSim uses packet-by-packet simulation model • Gives detailed simulation results • Significant computation costs • For large-scale networks with thousands of nodes, simulations are time consuming with discrete event simulators • Fluid-based analytical model uses differential equations to model network behavior
Network End nodes Project Goals • Integrate analytical model into GloMoSim to abstract out packet details • Possibility of speeding up simulation time and scaling well to large networks
Challenges • Interfacing between GloMoSim (discrete-event) and Fluid-flow model (analytical) • GloMoSim: transmission of packets, message exchanges, layered stack structure • Fluid-flow: differential equations, network characteristics obtained through formulas
Work Progress • UDP flows (no flow control) added to Fluid-based model • Fluid-based model converted from Matlab to C with wrapper functions • Model to be integrated at MAC layer (wiredlink model in GloMoSim) • Additional statistics data collected in GloMoSim
MATLAB Compiler Fluid-Flow Model (MATLAB) Fluid-Flow Model (C code) Function Wrapper (C code) Data for Fluid-flow States GloMoSim MAC layer (PARSEC) GloMoSim - Fluid Flow MAC layer Code Structure
Implementation Details • Original Fluid-flow model implemented in Matlab (by Misra et al.) • Inputs: network topology, capacity, initial window and queue sizes, number of flows, etc. • Outputs: TCP Window size for each flow, instantaneous and average queue lengths for each router as a function of time instants • Matlab code converted to C using Matlab compiler
Implementation Details • Each time fluid-flow analysis routine is invoked, it solves the differential equations for the given time period • FF Routine outputs queue size information of each router at the end of given execution time period. • Key: invoke FF routine with appropriate start and ending times, corresponding network topology and queue sizes
Simulation Topology • In GloMoSim: Single source, Single destination, one UDP Connection • Intermediate nodes in between modeled by Fluid-flow model (with background TCP) • GloMoSim FF side (source): convert packets entering FF model to flows • FF GloMoSim side (destination): monitor output data rate flowing out of FF model, then hand packets from source to destination
Simulation Topology • Two routers • 1 TCP (SgmsDgms) in GloMoSim, • 1 TCP (SffDff) within Fluid-flow model GMS node R1 R2 Sgms Dgms GMS node Sff Dff Fluid Flow section
Simulation Topology • Use GloMoSim source and destination nodes with TCP connections, while the Fluid-flow model simulates intermediate traffic at lower layers • Fluid-flow model implemented at MAC layer; no end-to-end flow control; flows do not react to congestion • Background TCP traffic inside Fluid-flow model can be added
QualNet results • Same topology • 5 classes of TCP flows • Tail-drop was used with 200 packet buffer capacity • RED analysis later
Future Directions • Validation of mixed simulator • Performance measurements • Interleaving analytical and discrete-event models within a simulation scenario • Scaling to large networks