Hello, everyone! In this concluding blog post, I’m excited to highlight the achievements in the field of Joint Power and Rate Control in User Space, and also offer insights into the future of this research and development endeavor. If this is your first encounter with my work, I strongly encourage you to explore the introductory blog posts from GSoC ’22 and GSoC ’23, as they provide a comprehensive overview of resource allocation in IEEE 802.11 networks.
GSoC ’22 blog posts: Introduction Mid-Term Final
GSoC ’23 blog posts: Introduction Mid-Term
Passive-Minstrel-HT in user space
After the first-half of the GSoC ’23 coding timeline, I did some more modifications to the passive user space Minstrel-HT such that it is more robust and also compatible with the new ORCA API. In my mid-term report, I conducted passive measurements on a link between a BananaPi with an MT7615 chip as the Access Point (AP) and a Xiaomi Redmi 4A Gigabit Edition with MT7621 as the Station (STA). However, it’s worth noting that this setup did not support aggregation in the transmission frames.
Since the kernel Minstrel-HT also consider real-time Aggregate MAC Protocol Data Unit (AMPDU) length to calculate the estimated throughput, it was crucial to test to test the behaviour of the user space Minstrel-HT (Py-Minstrel-HT) with its kernel counterpart in an aggregation context. To do this, the experiment setup was changed to involve two identical TP-Link WDR4900 routers, both equipped with ATH9K chips. One router operated as an Access Point (AP), and the other served as a Station (STA)
Initially, the rate selection between kernel Minstrel-HT and Py-Minstrel-HT for the pure ATH9k link with frame aggregation showed a significantly higher disparity compared to previous experiments on the MT76 link, where errors were consistently below 2%. Upon closer examination, it became evident that during the refactoring of Rateman and Py-Minstrel-HT, there had been a change in how AMPDU length was calculated, resulting in an incorrect calculation.
Furthermore, upon observing a consistently higher error rate associated with the maximum probability rate annotated at the end of the Multi-Rate Retry chain (MRR), I discovered a peculiarity in the kernel Minstrel-HT code. Specifically, when encountering initial attempt statistics for a rate that wasn’t used, the code assigned a previous average success probability of 0. This behavior occurred exclusively in situations where data rates had inherited their success probabilities from higher rates within the same group. It remains uncertain whether this aspect of the code logic was intentional or not.
To enhance the robustness of the measurements, I made adjustments to the execution of Passive Py-Minstrel-HT. Now, before factoring in transmission data, it follows a new sequence. Initially, it explicitly sets the lowest supported data rate, then resets the kernel rate statistics, and subsequently waits until the effect of the kernel rate statistics becomes evident in the API output trace. This revised approach ensures strict alignment between the transmission statistics in user space and those in kernel space right from the outset. As a result of these modifications, the rate selection between the kernel and user space Minstrel-HT is now identical.
In the following section, I present two comparison experiments conducted between kernel Minstrel-HT and Py-Minstrel-HT. The results are presented in terms of the number of errors, discrepancies in rate selection between the kernel and user space Minstrel-HT, and the percentage of errors observed at each Multi-Rate Retry (MRR) stage. These MRR stages, numbered from 0 to 3, are populated with rates ranked in descending order of their estimated throughput, with the final MRR stage reserved for the maximum probability rate to enhance robustness.
Experiment 1
MRR Setting info:
0 {'correct_instances': 19997, 'incorrect_instances': 0, 'percent_error': 0.0}
1 {'correct_instances': 19997, 'incorrect_instances': 0, 'percent_error': 0.0}
2 {'correct_instances': 19997, 'incorrect_instances': 0, 'percent_error': 0.0}
3 {'correct_instances': 19997, 'incorrect_instances': 0, 'percent_error': 0.0}
4 {'correct_instances': 19997, 'incorrect_instances': 0, 'percent_error': 0.0}
Experiment 2
MRR Setting info:
0 {'correct_instances': 19981, 'incorrect_instances': 0, 'percent_error': 0.0}
1 {'correct_instances': 19981, 'incorrect_instances': 0, 'percent_error': 0.0}
2 {'correct_instances': 19981, 'incorrect_instances': 0, 'percent_error': 0.0}
3 {'correct_instances': 19981, 'incorrect_instances': 0, 'percent_error': 0.0}
4 {'correct_instances': 19981, 'incorrect_instances': 0, 'percent_error': 0.0}
Given the congruent behavior observed between the user space and kernel space Minstrel-HT in terms of statistics collection and rate selection, this sets a robust foundation for expanding the capabilities of the user space Minstrel-HT with power control. This extension will allow for a performance comparison between the user space Minstrel-HT and kernel Minstrel-HT, offering valuable insights into their respective impacts.
Extending Minstrel-HT with Power Control (Joint Controller)
In this section, I will detail the implemented joint power and rate controller, which draws its inspiration from Minstrel-Blues, a framework originally developed by Prof. Thomas HΓΌhn in 2011. The joint controller has demonstrated its effectiveness by notably reducing interference and enhancing spatial reuse, particularly in scenarios involving multiple access points that utilize the joint controller. The subsequent sub-sections will present the algorithm in a format mirroring the structure of the proposed joint controller, facilitating a straightforward comparison between the two.
Configurable Parameters
The parameters listed in the table are the precise variable names used when specifying the rate control options (rc_opts) to Py-Minstrel-HT via Rateman.
Initialisation
During initialisation, the lowest supported rates are set using the reference power, unless Py-Minstrel-HT is executed with fixed power mode. Given that the implementation already incorporates a reference power, the ceiling mode has been eliminated from the extended controller.
Updating Rate Statistics
During each update interval, the extended Minstrel-HT algorithm updates the statistics for all the rates and power levels currently in use. Moreover, it also updates the success probability for unused rates at each power level, provided that the rate group contains at least one higher rate with attempt statistics from the same group and the same power level.
Following the statistics update, the selection of the reference power and sample power for all rates is adjusted based on their success probability and the constraints specified in the rc_opts parameters. If the success probability of the sample power is within the the inc_prob_tol of the reference power, the sample power is decrease by pwr_dec. On the other hand, if the success probability of the sample power is lower the the dec_prob_tol of the reference power, the sample power is increased by pwr_inc to increase throughput. The reference power is also updated in the same way but the tolerance being based on 100% (1.0) success probability.
For example, if the successful probability of the reference power is 0.95, the dec_prob_tol is 0.1 and the pwr_dec is 1, then the reference power will decrease by 1 dB because the probability of the reference power is greater than (1.0 – 0.1)=0.9. Finally, the best power level for all the rates is set to sample_power + opt_pwr_offset.
Rate and Power Sampling
The rate sampling process remains unchanged, as it continues to be carried out by the Minstrel-HT algorithm just as it was before the extension. However, there is a modification in that the power annotation for the sampled rates is now set to the reference power. Furthermore, the algorithm has been expanded to include power sampling, which cycles through the Multi-Rate Retry (MRR) selection process in a cyclic manner.
Selecting Best Rates for the MRR chain
In order to select the best rates, the algorithm defines a linear utility function that exposes trade-off between throughput (benefit) and interferences to other transmissions (cost) of each rate.
The benefit function is defined based on its estimated throughput and the current maximal throughput.
The cost factor serves as a representation of the interference costs associated with achieving a specific throughput at a given power level. In essence, this interference cost depends on two key factors: the extent of interference coverage and the duration of the interference event. However, due to the inherent complexity of modeling such effects, the following definition opts for a simplified approach, approximating the interference area solely based on power(ratei). Similarly, the duration of interference is approximated as 1 divided by the throughput (1/throughput(ratei)).
Using the utility function, the extended joint controller selects the best rates which have the highest utility, contrary to the Minstrel-HT algorithm which only considers the estimated throughput.
Analysing Performance of the Joint Controller
In this section, I will demonstrate the performance of the joint controller in comparison to the kernel Minstrel-HT and the Py-Minstrel-HT without power control extension.
Experiment Description
The experiments consisted of two variations based on the measurement tool:
Iperf3 and tcpdump
The traffic for the measurement was generated using iperf3, with the server hosted on the STA and the client on the AP. Consequently, the traffic flow traversed from the AP to the STA, aligning with the experiment’s objective of running resource controllers on the AP. Furthermore, the network traffic was monitored using tcpdump with a snapshot length of 150 bytes per packet to limit the amount of payload data captured. This data was then stored in a pcap file, which we subsequently parsed to extract the achieved throughput values.
The experiment setup involved two TP-Link WDR4900 routers, both featuring ATH9K chips. One router operated as an Access Point (AP), while the other functioned as a Station (STA).
Flent
Flent is a tool that serves as a wrapper around netperf and similar utilities to execute predefined tests, aggregate the outcomes, and generate plots. It also retains the raw data, providing users the opportunity to analyze and visualize the data, alongside the automated plots. The measurement traffic was generated using flent, with the server hosted on the AP and the client on the STA. However, the traffic direction was inverted using the “–swap-up-down” option. Notably, Flent can measure packet latency, a capability unavailable with iperf3.
The experiment setup involved a TP-Link WDR4900 router as an Access Point (AP) with a Macbook Pro 13′ 2017 as the Station (STA).
Results
The results have revealed a rather surprising trend: the joint controller consistently delivers improved throughput, even in single-link experiments. When compared to the kernel Minstrel-HT and Py-Minstrel-HT without the power control extension, the joint controller demonstrates a noteworthy throughput gain in UDP experiments, consistently falling within the range of 10-25%. These results highlight the joint controller’s capability to improve aggregation and dynamically select power levels.
Additionally, the results showcase various runs of the joint controller, each with a utility weight factor set to 1, 10, and 100. It is worth noting that the experiment using Flent utilised TCP traffic.
Iperf3 and tcpdump
Flent
Conclusion and Outlook
The analysis of the joint controller shows immense promise, and I’m excited about the prospect of continuing to run and test it with multiple interfering routers even beyond the conclusion of GSoC ’23. Regrettably, newer WiFi chips are becoming increasingly closed-source, limiting access to information and functionality related to resource control for the general public. Unfortunately, these newer chips still support power control and provide real-time estimates of throughput. As a result, our plan is to develop an independent power controller after completing Minstrel-Blues in the user space.
