Computational offloading, as an effective way to extend the capability of resource-limited edge devices in Internet of Things (IoT), is considered as a promising emerging paradigm for coping with delay-sensitive services. However, on one hand, applications commonly include several subtasks with dependent relations and on the other hand, the dynamic changes in network environments make offloading decision-making become a coupling and complex NP-hard problem, difficult to address. This paper proposes an intelligent Computational Offloading scheme for Dependent IoT Application ( <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">CODIA</monospace> ), which decouples the performance enhancement problem into two processes: scheduling and offloading. First, a prioritized scheduling strategy is designed and its complexity is analyzed. Then, an offloading algorithm with offline training and online deployment is introduced. Due to the temporal continuity between subtasks, the dependency relation is transformed into a transition of device state, and the overhead for the whole application is considered to be the long-term benefit. <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">CODIA</monospace> leverages an Actor-Critic-based solution, where the IoT devices are able to deploy intelligent models and dynamically adjust the offloading strategy to achieve low latency, while controlling energy consumption. Finally, a series of experiments are conducted to verify the robustness and efficiency of the proposed solution in terms of convergence, latency, and energy consumption. 

Edge Intelligence: A Computational Task Offloading Scheme for Dependent IoT Application

With the rapid development of Internet-of-Things (IoT) and the explosive advance of deep learning, there is an urgent need to enable deep learning inference on IoT devices in Mobile Edge Computing (MEC). To address the computation limitation of IoT devices in processing complex Deep Neural Networks (DNNs), computation offloading is proposed as a promising approach. Recently, partial computation offloading is developed to dynamically adjust task assignment strategy in different channel conditions for better performance. In this paper, we take advantage of intrinsic DNN computation characteristics and propose a novel Fused-Layer-based (FL-based) DNN model parallelism method to accelerate inference. The key idea is that a DNN layer can be converted to several smaller layers in order to increase partial computation offloading flexibility, and thus further create the better computation offloading solution. However, there is a trade-off between computation offloading flexibility as well as model parallelism overhead. Then, we investigate the optimal DNN model parallelism and the corresponding scheduling and offloading strategies in partial computation offloading. In particular, we propose a Particle Swarm Optimization with Minimizing Waiting (PSOMW) method, which explores and updates the FL strategy, path scheduling strategy, and path offloading strategy to reduce time complexity and avoid invalid solutions. Finally, we validate the effectiveness of the proposed method in commonly used DNNs. The results show that the proposed method can reduce the DNN inference time by an average of 12.75 times compared to the legacy No FL (NFL) algorithm, and is very close to the optimal solution achieved by the Brute Force (BF) algorithm with the difference of less than 0.04%.

Accelerating Deep Learning Inference via Model Parallelism and Partial Computation Offloading

With the rapid development of Internet-of-Things (IoT) and the explosive advance of deep learning, there is an urgent need to enable deep learning inference on IoT devices in Mobile Edge Computing (MEC). To address the computation limitation of IoT devices in processing complex Deep Neural Networks (DNNs), computation offloading is proposed as a promising approach. Recently, partial computation offloading is developed to dynamically adjust task assignment strategy in different channel conditions for better performance. In this paper, we take advantage of intrinsic DNN computation characteristics and propose a novel Fused-Layer-based (FL-based) DNN model parallelism method to accelerate inference. The key idea is that a DNN layer can be converted to several smaller layers in order to increase partial computation offloading flexibility, and thus further create the better computation offloading solution. However, there is a trade-off between computation offloading flexibility as well as model parallelism overhead. Then, we investigate the optimal DNN model parallelism and the corresponding scheduling and offloading strategies in partial computation offloading. In particular, we propose a Particle Swarm Optimization with Minimizing Waiting (PSOMW) method, which explores and updates the FL strategy, path scheduling strategy, and path offloading strategy to reduce time complexity and avoid invalid solutions. Finally, we validate the effectiveness of the proposed method in commonly used DNNs. The results show that the proposed method can reduce the DNN inference time by an average of 12.75 times compared to the legacy No FL (NFL) algorithm, and is very close to the optimal solution achieved by the Brute Force (BF) algorithm with the difference of less than 0.04%. 

Dependency-aware task offloading based on deep reinforcement learning in mobile edge computing networks

This paper considers computation offloading for mobile applications with task-dependency requirements in mobile edge computing (MEC) systems. Based on the online arrival patterns and various delay constraints of practical applications, we focus on minimizing the system deadline violation ratio (DVR) to improve the overall reliability performance. Specifically, we propose a DVR minimization computation offloading scheme with task migration and merging, in which the task migration and merging model is designed to construct an overall directed acyclic graph (DAG) for all currently dependent tasks. We consider a multi-slot MEC system where applications arrive slot-by-slot without prior knowledge of future arrivals. Then given the number of application arrivals at each time slot, we equivalently transform the DVR minimization problem into a problem that maximizes the number of completed applications in a finite time horizon. The above problem is challenging to determine the optimal task execution order for different applications with various task dependencies and delay constraints. To address this, we develop a migration-enabled multi-priority task sequencing algorithm, which creatively introduces several task priority metrics and determines the optimal task execution order. Then, a deep deterministic policy gradient (DDPG)-based learning algorithm is developed to find the optimal offloading policy. Experimental results demonstrate that the proposed scheme can reduce the system DVR by 60.34%~70.3% compared with existing benchmark schemes under various network scenarios. 

Dependent Task Scheduling and Offloading for Minimizing Deadline Violation Ratio in Mobile Edge Computing Networks

Computation offloading is one of the main use-cases of the Multi-access Edge Computing (MEC) paradigm which can help to save the battery life of the resource-poor mobile devices by transferring the computation-intensive tasks to the resource-rich edge cloud servers. However, the ever-increasing internet traffic can negate this benefit due to possible failures of the MEC servers. On the other hand, the growth of computation capabilities of end devices as the result of recent developments of Central Processing Units (CPUs), can help to enhance the MEC systems performance and broaden the concept of edge computing by using device to device communication (D2D). The majority of existing works on computation offloading assume the tasks are independent and can be executed in parallel; however, the dependency among tasks can introduce new problems. To investigate this issue, in this paper, we consider a basic three-node MEC system consists of a user node with sequentially-dependent tasks, a helper/relay node, and a MEC server located at the base station (BS). We formulate the offloading problem into an energy-efficiency minimization problem while satisfying the task-dependency and completion deadline requirements. The simulation results show the superior performance of our proposed method compared to the other approaches.

Energy-Aware Cooperative Offloading Framework for Inter-dependent and Delay-sensitive Tasks

We present a high-stability regulation circuit to ensure the safety of a device within a wide range of the back-sink current for a USB Type-C interface application. The proposed adaptive linear pole–zero tracking compensation can linearly compensate for the changes in the back-sink current, thereby adaptively canceling the pole–zero changes caused by the current changes. The simulation results show that the phase margin remains greater than 60°. Meanwhile, the loop bandwidth changes between 45 kHz and 135 kHz, when the current increases from 0 A to 1 A, ensuring excellent loop stability. The high-stability regulation circuit is realized in a standard 180 nm CMOS process with an area of 0.4 mm × 0.6 mm. The chip regulates an output voltage from 4.5 V to 5.5 V with 1 A current capacity and 100 mV maximum dropout voltage with the help of the adaptive linear pole–zero tracking compensation.

/pdf/a-high-stability-regulation-circuit-with-adaptive-linear-3blvmqcu.pdf

Yuanfei Wang

Papers

Energy-Aware Cooperative Offloading Framework for Inter-dependent and Delay-sensitive Tasks

A High-Stability Regulation Circuit with Adaptive Linear Pole–Zero Tracking Compensation for USB Type-C Interface