PI Koutsonikolas’ research for the past one year focused on the experimental exploration of the two primary link adaptation mechanisms in 60 GHz WLANs, namely rate adaptation (RA) and beam adaptation (BA). The IEEE 802.11ad/ay standards mandate the use of both mechanisms but they do not specify when each of the two adaptation mechanisms should be used or in what order, and 60 GHz chipset vendors resort to heuristics to select the right mechanism, which often fail to make the right choice even in seemingly simple scenarios. To fill this gap, PI Koutsonikolas designed LiBRA (Learning-based Beam and Rate Adaptation), a practical, standard-compliant, learning-based link adaptation framework that leverages PHY layer information to determine (i) when to trigger link adaptation and (ii) which of the two adaptation mechanisms to trigger, and works with a variety of RA and BA algorithms.
Using a variety of commercial off-the-shelf 802.11ad devices, we evaluate the effectiveness of the heuristics such devices used to determine when to trigger each of the two adaptation mechanisms. We explore for first time the feasibility of utilizing PHY layer information to guide link adaptation in 60 GHz WLANs using a large dataset (to be made publicly available) collected with the X60 testbed in a variety of indoor environments and scenarios. We investigate the effectiveness of a number of PHY layer metrics in predicting which of the two mechanisms should be triggered at a given scenario. We also explore for first time ML-based link adaptation approaches leveraging PHY layer information.
We design LiBRA, a practical, standard-compliant, learning-based link adaptation framework that leverages PHY layer information to determine (i) when to trigger link adaptation and (ii) which of the two adaptation mechanisms to trigger, and works with a variety of RA and BA algorithms. LiBRA strikes a balance between two performance metrics – throughput and link recovery delay. We evaluate LiBRA using extensive trace-based simulations with different sets of realistic PHY and MAC layer parameters.
Our experiments with commercial off-the-shelf devices showed that simple heuristics used by such devices often fail to select the right adaptation mechanism (RA or BA) even in seemingly simple scenarios. Such wrong decisions can lead to significant performance degradation.Using the X60 dataset, we found that, while some PHY layer metrics turn out to be more useful than others, no metric works in all scenarios, suggesting that a combination of metrics is required, and motivating the use of ML-based approaches. Using the same dataset and 5-fold cross-validation, we showed that simple models based on random forests can predict the right action with 98% accuracy. We further evaluate the models on a new dataset collected from two different buildings showing that they retain satisfactory accuracy.
We evaluated LiBRA using extensive trace-based simulations with different sets of realistic PHY and MAC layer parameters (aggregated frame size, beam searching duration, flow duration) in scenarios involving various link impairments — linear and/or angular displacement, motion, human blockage, interference, and a combination of them. Our results show that LiBRA performs closely to an oracle that always selects the right adaptation mechanism depending on the performance metric we want to optimize (amount of downloaded data or link recovery delay) and significantly outperforms two simple heuristics used by commercial off-the-shelf devices.
Publications
Shivang Aggarwal, Urjit Satish Sardesai, Viral Sinha, and Dimitrios Koutsonikolas, “Poster: Link Adaptation in 60 GHz WLANs using PHY Layer Information”, in the 21st ACM International Workshop on Mobile Computing Systems and Applications (HotMobile), Austin, TX, March 3-4, 2020.
Shivang Aggarwal, Urjit Satish Sardesai, Viral Sinha, Deen Dayal Mohan, Moinak Ghoshal, and Dimitrios Koutsonikolas, “LiBRA: Learning-Based Link Adaptation Leveraging PHY Layer Information in 60 GHz WLANs”, under submission.