Abstract

In discrete-event network simulations, a significant portion of resources are dedicated to the processing of packet events. For large-scale network simulations, the processing of packet events is the most time consuming aspect of the simulation. In this work, we develop a technique that saves on the processing of packet events for TCP flows using the well established results showing that the average behavior of a TCP flow is predictable given a steady-state path condition. We exploit this to predict the average behavior of a TCP flow over a future period of time where steady-state conditions hold, thus allowing for a reduction (or elimination) of the processing required for packet events. We call the flows simulated with our scheme “predicted flows”, in contrast to the “packet-based flows” in usual packet level simulations. We design a simulation framework that provides the flexibility to incorporate this method of simulating TCP packet flows. Our goal is (1) to accommodate different network configurations, on/off flow behaviors and interaction between predicted flows and packet-based flows; and (2) to preserve the statistical behavior of every entity in the system so as to maintain the accuracy of the network simulation as a whole. In order to illustrate the promise of this idea, we implement it in the context of the ns2 simulation system. A set of experiments illustrate the speedup and approximation of the simulation framework, and relate the effectiveness of the scheme to the transient behavior of TCP congestion control. We compare two approaches to predicting TCP throughput and show that formula-based prediction suffers inaccuracy under certain scenarios due to the simple assumptions made about inter-loss time, as most fluid models do, but otherwise achieves a good approximation faster than direct measurement prediction.

Keywords: TCP throughput equation; TCP modeling; Event-based simulation; Fluid model

Article Outline

1. Introduction

2. Prediction-based simulation design

2.1. Simulation context

2.2. Simulation framework details

2.2.1. Epochs

2.2.2. Intra-epoch transition

2.2.3. Inter-epoch transition

3. Predicting steady-state throughput

3.1. Summary of TCP behavior

3.2. Measurement period

3.3. Direct throughput measurement

3.4. Formula based prediction

4. Mixing packet-based and predicted flows

4.1. Changes to the framework

4.2. Accounting for network resources usage

5. Experimental results

5.1. Implementation

5.2. Speedup

5.3. Approximation

5.3.1. Average throughput

Impact of loss ratio and RTT

Impact of queuing discipline

Impact of the number of interacting flows

Bandwidth and delay heterogeneity

5.3.2. Buffer occupancy

5.3.3. Mixing packet-based and predicted flows

Packet sequence traces

Buffer occupancy and loss rate

5.4. Experience with larger IP networks

5.4.1. Approximation

5.4.2. Speedup

6. Suggestions for usage

6.1. Applicability

7. Comparison with fluid flow models

8. Concluding remarks

References

Vitae