DELTA: Deadline aware energy and latency-optimized task offloading and resource allocation in GPU-enabled, PiM-enabled distributed heterogeneous MEC architecture

Abstract

The use of Multi-access Edge Computing (MEC) technology holds great potential for supporting modern, computation-intensive, and time-sensitive applications. These applications are mainly generated from resource-constrained handheld or mobile user equipment (UE). As these devices have limited resources and some are also energy-constrained, it is crucial to offload some portions of the applications (or tasks) to the connected MEC servers. However, MEC servers also have limited resources compared to cloud servers, making it imperative to implement efficient task-offloading policies for UE devices and optimal resource allocation policies for MEC servers. In this paper, we first formulate the energy and latency minimization problem as a multi-objective Mixed Integer Programming (MIP) problem, and we propose a novel deadline-aware energy and latency-optimized task offloading and resource allocation (DELTA) strategy to execute the applications on a cooperative heterogeneous MEC architecture efficiently. Our policy aims to minimize the energy consumption of UEs and the latency of applications while meeting the deadline and dependency constraints of the applications. In our heterogeneous cooperative MEC system, as a novel contribution, we consider that the MEC servers are equipped with graphics processing unit (GPUs), solid-state disk (SSD) storage, and processing in-memory (PiM) enabled memory, in addition to the traditional processors, memories, and hard disk storage. Furthermore, we consider the UEs to be dynamic voltage and frequency scaling (DVFS) enabled. We perform an extensive simulation using the real data set on a standard simulator and compare our results with three different policies (Intelligent-TO (Chen et al., 2023), Multi-user (Yang et al., 2020) and Selective-random). Our proposed strategy DELTA achieves a 71.18% reduction in latency on average compared to the considered state-of-the-art policy, and it outperforms the most efficient benchmarked strategy, Intelligent-TO, by 59.6% in terms of latency. Regarding energy consumption for UE devices, the considered state-of-the-art policies consume about 4x more energy on average than the DELTA. Although Intelligent-TO is the most energy-efficient policy among those benchmarked, DELTA surpasses it, achieving a 59.6% reduction in energy consumption.