Quality-of-Service Driven Resource Allocation Based on Martingale Theory

One of the key metrics in measuring system quality of service (QoS) is the delay performance. Most existing papers have focused on the studies of decreasing transmission delay. However, as the wireless communication traffic increasing dramatically, queueing delay in the wireless networks becomes a non-negligible issue. Martingale theory, which fits any arrival and service process, providing a much tighter delay bound compared to the effective bandwidth theory, has been proposed to analyze the system queueing delay bound, especially in a bursty traffic scenario. In this paper, we propose to study the resource allocation problem based on the delay bounds derived from martingale theory. In specific, we first revisit some basic knowledge about stochastic network calculus, and present the delay bounds derived from martingale theory in certain typical bursty service models. Then, we setup a resource allocation problem in a computation offloading scenario, where multiple computation nodes with distinct computation capacities are considered. User's computation tasks are usually bursty, and are required to be executed within a limited time. We propose to minimize the system delay violation probability by properly allocating the computation tasks to different computation nodes. A closed-form solution is derived for the computation offloading problem, using a special kind of water-filling policy. Moreover, we discuss two potential models of martingale-based resource allocation, and provide the corresponding system architectures. Finally, numerical results are presented to demonstrate the performances of the proposed scheme. The proposed water-filling scheme achieves a smaller system delay violation probability compared to the benchmark.