Practical Machine Learning-Based Rate Adaptation Solution for Wi-Fi NICs: IEEE 802.11ac as a Case Study

Abstract

Many rate adaptation (RA) solutions have been proposed for legacy Wi-Fi standards. However, these solutions lack extensibility, and cannot therefore be easily applied to new Wi-Fi standards such as IEEE 802.11ac. Furthermore, while several practical RA solutions have been proposed for IEEE 802.11ac network interface cards (NICs), e.g., Minstrel and Iwlwifi, they achieve poor goodput performance in some cases since they lack scalability or thoroughness. Given a large number of rates in IEEE 802.11ac, the lack of scalability results in a slow convergence on the search of the best rate over time. To accelerate the rate search, several RA solutions skip some rates to narrow down the scope of candidate rates, but lack thoroughness and may miss the best rate. Accordingly, the present study proposes a holistic RA solution, designated as Machine Learning-based RA (MLRA), which is not only practical, but also extensible, scalable and thorough. MLRA achieves the extensible property by leveraging machine learning to automatically identify the correlations among the rate, goodput performance and link quality. Moreover, it achieves the scalable and thorough properties using a two-level rate search procedure and a congestion detector. The performance of MLRA is evaluated on commodity IEEE 802.11ac NICs using TensorFlow with an asynchronous framework implemented across the kernel and user spaces of the operating system. The results show that MLRA outperforms other practical RAs by up to around 133.0%-658.0% with a negligible overhead.