Scheduling wireless transmissions exploiting application awareness and knowledge of the future
This dissertation explores ways to improve the scheduling of wireless transmissions, by exploiting the application layer information of the ongoing transmissions and exploiting the knowledge of the future capacity variations. First, we consider the design of cross-layer opportunistic transport protocols for stored video over wireless networks with a slow varying (average) capacity. We focus on two key principles: (1) scheduling data transmissions when capacity is high; and (2), exploiting knowledge of future capacity variations. The latter is possible when users' mobility is known or predictable, e.g., users riding on public transportation or using navigation systems. We consider the design of cross-layer transmission schedules which minimize system utilization (and thus possibly transmit/receive energy) while avoiding, if at all possible, rebuffering/delays, in several scenarios. For the single-user anticipative case where all future capacity variations are known beforehand; we establish the optimal transmission schedule is a Generalized Piecewise Constant Thresholding (GPCT) scheme. For the single-user partially anticipative case where only a finite window of future capacity variations is known, we propose an online policy: Greedy Fixed Horizon Control (GFHC). An upper bound on the competitive ratio of GFHC and GPCT is established showing how performance loss depends on the window size, receiver playback buffer, and capacity variability. We also consider the multiuser case where one can exploit both future temporal and multiuser diversity. Finally we investigate the impact of uncertainty in knowledge of future capacity variations, and propose an offline approach as well as an online algorithm to deal with such uncertainty. Our simulations and evaluation based on a measured wireless capacity trace exhibit robust potential gains for our proposed transmission schemes.
Second, we consider the design of scheduling algorithms for dynamic D2D networks where links are set up to mediate file transfers among close by users, but with limited `contact' times or deadlines. Our focus is on improving three performance metrics: (1) overall offload data; (2) number of transfers which complete within their deadlines; and (3) file transfer delays. The design and evaluation of existing D2D scheduling algorithms has mostly focused on optimizing the network's sum rate subject to fairness concerns for a fixed set of links. Starting from a dynamic scenario where links share a single collision domain, this dissertation investigates optimal scheduling algorithms which exploit application layer context. These results drive our proposal for application-aware versions of state-of-the-art schedulers such as FlashLinQ and CSMA/CA. Our simulations show that these schedulers could achieve substantial performance improvements depending on the operational scenario, i.e., density of users, file and contact time distributions, etc. We also show that performance can be further enhanced by incorporating both application-aware scheduling and admission control. Finally we investigate scenarios where D2D links have no deadlines. We show that application-aware schedulers result in smaller average file transfer delays and can increase the stability region of the system. This is in part due to their ability to reduce spatial clustering of links resulting from interference coupling in the dynamic setting.