Measurement-Based Opportunistic Scheduling for Heterogenous Wireless Systems

We study the performance of an opportunistic scheduling scheme
maximum quantile scheduling
, i.e., scheduling a user whose current rate is in the highest quantile relative to its current rate distribution, in a wireless system. In a practical scenario it is unlikely that users' rate distributions are known at the scheduler, and have to be estimated via measurement. Under the assumption of fast fading, we prove a bound on the relative penalty associated with such estimates, showing that number of independent samples need only grow linearly with the
number of active users
. This is a fairly limited cost, suggesting one could track distributional changes in users' channels. By contrast other opportunistic scheduling schemes require estimating or setting weights/thresholds that implicitly depend not only on the number of users, but also
their rate distributions
, and
possibly their traffic characteristics
. In other words the penalty associated with tuning weights for other schemes can be higher than that associated with estimating users' rate distributions for maximum quantile scheduling. This statement is supported by our simulation results. Furthermore we prove that if rates are bounded and number of users is high, maximum quantile scheduling is sum average throughput maximizing subject to temporal fairness.