Performance Evaluation. of Input and Virtual Output Queuing on. Self-Similar Traffic

Transcription

1 Page 1 of 11 CS 678 Topics in Internet Research Progress Report Performance Evaluation of Input and Virtual Output Queuing on Self-Similar Traffic Submitted to Zartash Afzal Uzmi By : Group # 3 Muhammad Saeed Syed Imran Jami

2 Page 2 of 11 Table of contents 1 Abstract 2 Introduction 2.1 Self-Similar Traffic 2.2 Self-Similar Traffic Distribution 2.3 Network Simulator Architecture 3 Simulation Model 3.1 Assumptions General Assumptions Input Queuing Virtual Output Queuing 3.2 Implementation 4 Conclusion 5 Reference 6 Appendix

3 Page 3 of 11 Abstract The work of [MKMG87] and [NMJW] on input and virtual output queue simulation were based on Poisson distribution traffic. However the work of [CROV97] showed that the real traffic is of self similar nature. In this work, we are performing the simulation, by using the model of [MKMG87] and [NMJW], on Poisson and Self-Similar traffic, and then compare their results to check their validity. This document shows the progress of our research till mid of the quarter. It first shows about our study of self similar traffic distribution and then the tool we will use for simulation. The second part discusses about the implementation. It deals with the model of simulation and its implementation in NS Introduction

4 Page 4 of Self-Similar Traffic A self-similar phenomenon displays structural similarities across a wide range of timescales. Selfsimilarity is the property associated with "fractals," which are objects whose appearances are unchanged regardless of the scale at which they are viewed [CROV97]. In the case of stochastic objects like time series, self-similarity is used in the distributed sense. When viewed at varying timescales, the object's relational structure remains unchanged. As a result, such a time series exhibit bursts at a wide range of timescales [SHAI99]. Thus traffic that is bursty on many or all timescales can be described statistically using the notion of self-similarity. 2.2 Self-Similar Traffic Distribution In the work of [VPSF95], it has been pointed out, that self-similar network traffic can be generated with the help of ON-OFF sources. An ON-OFF source alternates between sending packets (ON) and being idle (OFF). If the durations of the ON and the OFF phases are described by heavy tailed probability distributions then the superposition of a large number of ON-OFF sources result in self-similar packet traffic [WTSW97]. Heavy tailed distributions have high or even infinite variance and therefore show extreme variability on all time scales. The simplest type of heavy-tailed distribution is the Pareto distribution. Study done by many researches show that Ethernet, WWW and FTP data sessions follow Pareto distribution. Research done by [CROV97] has revealed that the traffic generated by World Wide Web traffic shows self-similar characteristics. Comparing the distributions of ON and OFF times, they found that the ON time distribution was heavier-tailed than the OFF time distribution. The distribution of file sizes in the Web might be the primary determiner of Web traffic self-similarity. Research done by [VPSF95], showed that in FTP data sessions have burstier arrival rates, and number of bytes in each burst has a heavy upper tail. [WTSW97] showed that a Pareto heavy-tailed distribution reflects the actual behavior of an individual Ethernet source.

5 Page 5 of Network Simulator Network Simulator (NS 2) is an object-oriented, discrete event driven network simulator. It is useful in the simulation of networks based on LAN and WAN by using protocols such as TCP and UDP. The traffic source behavior can be based on FTP, Telnet, Web, CBR and VBR. The scheduling algorithms for queue in NS are Drop Tail, RED and CBQ. NS also implements multicasting and some of the MAC layer protocols for LAN simulations Architecture In NS, user first writes a Tcl script on any editor, which is send to the Tcl interpreter. The Tcl interpreter has the following components. Event scheduler Network component object libraries and Network setup module libraries The Tcl interpreter initiates an event scheduler, sets up the network topology using the network objects and functions in the library, and tells traffic sources when to start and stop transmitting packets through the event scheduler. In order to obtain simulation results, NS produces text-based output files that contain detailed simulation data. The data can be used for simulation analysis on Excel or any other spreadsheet package. If indicated in script, simulation data can also be provided graphically on a graphical simulation display tool called Network Animator (NAM). NAM can also present information about throughput and number of packet drops at each link. Xgraph and other GUI based tools can also be used for this purpose. 3. Simulation Model

6 Page 6 of 11 We are using the same model for input and virtual output queuing as used in [MKMG87] and [NMJW], for both Poisson and Self-Similar traffic. The main difference is in probability distribution for inter arrival times. For Poisson distribution we are generating traffic with Exponential ON/OFF distribution, and for self-similar traffic (heavy-tailed) we are using Pareto ON/OFF distribution. 3.1 Assumptions Our simulations model consists of following assumptions General Assumptions Non-Blocking Switches We are considering only non-blocking switch that allows establishment of connection at any time between any pair of idle input-output ports. Crossbar switch is an example, with N 2 inter-connections of switch fabric Fixed-Length Packet Switch operates synchronously with fixed-length packets. Packet Scheduling During any time slot, packet may arrive at any input port which will then be destined to any output port. As we are maintaining queues at the input ports, switch fabric will operate at the same speed as the input/output ports. During each time slot only one packet may be accepted for any given output and remaining packets will be queued at the input queue. 3.2 Input Queuing In our model, the assumption is made that packet arrivals on the N input will be independent and identical Bernoulli processes. Suppose that the probability that a packet will arrive on a particular input is p, then the probability of each packet handled by corresponding output is 1 /N. If several packets addressed to the same output port at a single time slot then one of the packet will be selected randomly or from a longest queue whichever is simple to implement in NS. 3.3 Virtual Output Queuing The model consists of a switch with M inputs and N outputs. The arrival process is assumed to be stationary and ergodic. The set of all arrival processes is considered Admissible if no input or output is oversubscribed otherwise it is Inadmissible. An edge joining non-empty input queues with the destined output queues makes a bipartite graph.

7 Page 7 of 11 The FIFO queues are served with scheduling algorithms that selects a match/matching M between the input and outputs by solving the graph. There are two types of matching algorithms available. Maximum Size Matching Algorithm (finds the match containing the maximum number of edges/packet-transfer). Maximum Weight Matching Algorithm (finds the maximum weight, consider only integer-valued weights equaling the occupancy of each queue). We are using maximum weight matching algorithm in our simulation. 3.2 Implementation The project requires the simulation on two traffic generators: Poisson and Pareto (Self Similar) distribution. In simulating both traffic distribution on VOQ and input queued switch we identified the following components in our model: Set of nodes Links Traffic source Queue selection policy Traffic distribution The simulation will be carried out on a switch that consists of input and output ports that will be represented by nodes. In Tcl script nodes are added in a topology as: set nx [$ns node] where X is the number of node. The ports will be connected to each other by a switching fabric which is represented as a set of links connecting the nodes, with either simplex or duplex link. This is represented in Tcl script as: $ns duplex-link $nx 1 $nx 2 YMb Zms SP This line notifies the simulator object to connect two nodes with a duplex link (for example) with Y Mb bandwidth and Z ms delay. SP shows the scheduling policy of input queue (for example DropTail queue). Traffic source that we will use in this project is TCP. In NS 2, a TCP agent is added on input port which is then connected with Pareto traffic generator. In Tcl, an abstract view is: set tcpx [new Agent/TCP] $ns attach-agent $nx $udpx In generating self similar traffic a new Pareto On/Off traffic generator will be created as follows:

8 Page 8 of 11 Set p [new Application/Traffic/Pareto] $p set packetsize_ xyz $p set burst_time_ y ms $p set idle_time_ z ms $p set rate_ xy k $p set shape_ p Similarly, in generating traffic based on Poisson distribution, a new Exponential On/Off traffic generator will be created as follows: set q [new Application/Traffic/Exponential] $q set packetsize_ xyz $q set burst_time_ 0 ms $q set idle_time_ z ms $q set rate_ (very large value) k If burst time is set as 0 ms and rate of traffic is very large then Exponential distribution behaves like Pois $p (q) attach-agent $tcpx Still there are lot of study requires in NS 2 to do advance work, but we think that till now we will be able to do the required tasks successfully. 4. Conclusion Self Similarity is a phenomenon that displays structural similarities across a wide range of timescales. We discuss this issue in great detail and study NS 2 in order to model and analyze this behavior. Till midterm our tasks required to perform were NS study and Simulation model designing. Both of the tasks are completed successfully. After next week we will move on to our next task which is Tcl script writing. Study of advance features of NS 2 are still undergoing and by the end of this

9 Page 9 of 11 quarter we will InshAllah submit the required research paper. 5. References [CROV97] M. E. Crovella, "Self-Similarity in WWW Traffic: Evidence and Possible Causes", IEEE/ACM Transaction on Networking, December 1997 [SHAI99] Z. Sahinoglu and S. Tekinay, "On Multimedia Networks: Self-Similar Traffic and Network Performance", IEEE Communications Magazine, January 1999 [VPSF95] Verne Paxson and Sally Flyd. Wide area traffic: The failure of Poisson modeling. IEEE/ACM Transactions on Networking, [WTSW97] Walter Willinger, Murad S. Taqqu, Robert Sherman, and Daniel V. Wilson. Self-similarity Through High-variability: Statistical Analysis of Ethernet LAN Traffic at the Source Level. IEEE/ACM Transactions on Networking, February [WILL02] William Stallings, High-Speed Network and Internets, Performance and quality of service, 2 nd edition, [MKMG87] Mark J. Karol and Michael G. Hluchyj, Input Versus Output Queuing on Space Division Packet Switch, IEEE Transaction On Communication, December [NMJW] Nick McKeown and Jean Walrand, Achieving 100% Throughput in Input Queued Switch. [JCMC] Jae Chung and Mark Claypool, NS by Example, WPI Computer Science [MARC] Marc Greis, NS Tutorial, VINT group

10 Page 10 of 11 [LAVB] Lavon B. Page, Probability for Engineering, 1989 CS press. 6. Appendix Independent Process/Random Variable When the probability of occurring one event is independent of the occurrence of other event then these event are know as independent events. Bernoulli Random Variable/Processes A R.V or process is known to be a Bernoulli if there are only two possible outcomes either success or failure. We can also represent these values as 0/1 or ON/OFF. Stochastic Process Stochastic process is a random variable that is a function of time. Stationary Stochastic Process In which probability characteristics of process do not vary as a function of time. If its expected value is a constant and its autocorrelation function depends only on the time difference. Ergodic Process Stationary stochastic process is called ergodic if its Because Time average = ensemble average E[X(t)] is constant for stationary process E[Mt] = E[X(t)] = µ For a ergodic process Lim T 0 Var [Mt] = 0; Ensemble averages of SP In random variable set of all possible outcomes is called ensemble and have mean and variance for this called ensemble averages.

11 Page 11 of 11 Time averages We consider only a single outcome of X(t), random variable of expected values.