In the past decade, researchers have proposed a plethora of routing metrics to replace the traditional hopcount (HC) metric in legacy 802.11a/b/g WMNs. These metrics estimate the cost of a path based on the underlying link quality, taking into account factors such as link loss rates or link bandwidth which they measure using periodic probe packets. Our experimental studies in an 802.11n wireless mesh network testbed (UBMesh) revealed that the throughput gains of state-of-the-art link quality-based routing metrics over the HC metric in legacy WMNs do not carry over in 802.11n WMNs. Even worse, the large link throughput gains of 802.11n over 802.11a/b/g do not translate into multihop throughput gains. We found that popular probing techniques for link loss rate and bandwidth estimation yield poor accuracy in 802.11n WMNs due to the new features introduced at the underlying MAC/PHY layers. The problem is expected to deteriorate further in 802.11ac WMNs with a much larger number of MAC/PHY configurations -- MCS/channel width/MIMO stream combinations -- compared to 802.11n.
Since probing all available configurations is practically impossible, our current research aims at designing techniques for reducing the probing space. We are currently using machine learning-based techniques for identifying similar loss rate patterns among different configurations and developing techniques for learning such patterns online. Identifying such patterns will allow wireless mesh routers to create clusters of configurations with similar loss rates and only probe one configuration at each cluster.