A Hardware Testbed for Measuring IEEE 802.11g DCF Performance

The Distributed Coordination Function (DCF) is the oldest and most widely-used IEEE 802.11 contention-based channel access control protocol. DCF adds a significant amount of overhead in the form of preambles, frame headers, randomised binary exponential back-off and inter-frame spaces. Having accurate and verified performance models for DCF is thus integral to understanding the performance of IEEE 802.11 as a whole. In this document DCF performance is measured subject to two different workload models using an IEEE 802.11g test bed. Bianchi proposed the first accurate analytic model for measuring the performance of DCF. The model calculates normalised aggregate throughput as a function of the number of stations contending for channel access. The model also makes a number of assumptions about the system, including saturation conditions (all stations have a fixed-length packet to send at all times), full-connectivity between stations, constant collision probability and perfect channel conditions. Many authors have extended Bianchi's machine model to correct certain inconsistencies with the standard, while very few have considered alternative workload models. Owing to the complexities associated with prototyping, most models are verified against simulations and not experimentally using a test bed. In addition to a saturation model we considered a more realistic workload model representing wireless Internet traffic. Producing a stochastic model for such a workload was a challenging task, as usage patterns change significantly between users and over time. We implemented and compared two Markov Arrival Processes (MAPs) for packet arrivals at each client - a Discrete-time Batch Markovian Arrival Process (D-BMAP) and a modified Hierarchical Markov Modulated Poisson Process (H-MMPP). Both models had parameters drawn from the same wireless trace data. It was found that, while the latter model exhibits better Long Range Dependency at the network level, the former represented traces more accurately at the client-level, which made it more appropriate for the test bed experiments. A nine station IEEE 802.11 test bed was constructed to measure the real world performance of the DCF protocol experimentally. The stations used IEEE 802.11g cards based on the Atheros AR5212 chipset and ran a custom Linux distribution. The test bed was moved to a remote location where there was no measured risk of interference from neighbouring radio transmitters in the same band. The DCF machine model was fixed and normalised aggregate throughput was measured for one through to eight contending stations, subject to (i) saturation with fixed packet length equal to 1000 bytes, and (ii) the D-BMAP workload model for wireless Internet traffic. Control messages were forwarded on a separate wired backbone network so that they did not interfere with the experiments. Analytic solver software was written to calculate numerical solutions for thee popular analytic models for DCF and compared the solutions to the saturation test bed experiments. Although the normalised aggregate throughput trends were the same, it was found that as the number of contending stations increases, so the measured aggregate DCF performance diverged from all three analytic model's predictions; for every station added to the network normalised aggregate throughput was measured lower than analytically predicted. We conclude that some property of the test bed was not captured by the simulation software used to verify the analytic models. The D-BMAP experiments yielded a significantly lower normalised aggregate throughput than the saturation experiments, which is a clear result of channel underutilisation. Although this is a simple result, it highlights the importance of the traffic model on network performance. Normalised aggregate throughput appeared to scale more linearly when compared to the RTS/CTS access mechanism, but no firm conclusion could be drawn at 95% confidence. We conclude further that, although normalised aggregate throughput is appropriate for describing overall channel utilisation in the steady state, jitter, response time and error rate are more important performance metrics in the case of bursty traffic