Flow-level modeling and optimization of intercell coordination with dynamic TDD
Proceedings of the 10th ACM symposium on QoS and security for wireless and mobile networks - Q2SWinet '14
We study the intercell coordination problem between two interfering cells combined with dynamic time-division duplexing (TDD). In dynamic TDD, each station selects in each time slot whether it is serving uplink (u) or downlink (d) traffic. Thus, the system has four possible operation modes (uu, ud, du, dd). The amount of intercell interference between the stations clearly depends on the operation mode. We consider a flow-level model where traffic consists of elastic data flows in both cells
... ls 1 and 2) and in both directions (uplink and downlink). We first characterize the maximal stability region, and then determine the optimal static (i.e., state-independent) policy. Our main objective is to analyze the potential gains from applying dynamic (i.e., state-dependent) policies, where the chosen operation mode depends on the instantaneous state of the system. To this end, motivated by certain stochastic optimality results in the literature, we define several priority policies. As a reference policy, we have the well-known max-weight policy, and we also develop another dynamic policy by applying the policy iteration algorithm. Notably we prove that certain simple priority policies are, in fact, stochastically optimal in some special cases, but which policy is optimal depends on the setting. To study the exact performance gains achieved by the dynamic policies, we perform extensive simulations. While our stochastic optimality results require exponential service times, in the simulations, we also study the impact of nonexponential service times and consider a physical model where the service time distribution is determined by the joint distribution of flow sizes and the random location of the corresponding user in the cell area. The max-weight policy is, as expected, performing well but the various priority policies are sometimes better and even optimal. Jointly the results indicate that dynamic policies give significant performance gains compared with the optimal static policy.