Dynamic Offloading for Compute Adaptive Jobs

Agrim Bari, Gustavo de Veciana, Kerstin Johnsson, Alexander Pyattaev

Abstract: The increasing demands of computationally intensive device applications are driving advancements in edge technologies and the need for improved computation offloading policies. This paper focuses on “adaptive” offloading and computation, i.e., adapting the amount of offload data and, consequently, the associated compute adaptive job's quality to the wireless channel quality, network congestion, and the compute adaptive job's computational options. For example, when the channel quality is poor and a job has a tight deadline, the amount of offloaded data can be reduced in exchange for a loss in the associated compute adaptive job's quality. In this paper, we show the substantial advantages of adapting the amount of offloaded data to channel quality and network congestion. We begin by defining a reward model for compute adaptive jobs based on the amount of offloaded data and resulting computation quality. We then develop an upper bound on the achievable revenue rate and propose/compare various offloading policies: Greedy, Predictive Abandonment (PA), Probabilistic Admission Control and Layer Assignment (PACLA), and combinations thereof. We evaluate our policies via simulation and observe that the combination of PACLA + PA provides the best offloading performance for homogeneous or heterogeneous compute adaptive jobs for devices with varied channel qualities.