Task Delay Optimization For Multi-node Collaborative Computing In Edge Distributed Scenarios

As a novel computing paradigm,edge computing can alleviate the imbalance between the huge computing demand at the network edge and the limited computing power of the terminal.Since edge computing pushes cloud services from the network core to the network edge,it can greatly reduce the transmission delay of task data.However,the transmission delay advantage brought by the architecture alone can hardly meet the requirements of delay-sensitive applications.To further improve the overall efficiency,it is necessary to optimize the task delay for edge computing.This dissertation focuses on the edge distributed scenarios where multiple users or edge nodes coexist,and aims to optimize the delay performance for the multi-node collaborative computing tasks.Starting from the collaborative mechanism in the multi-user collaborative computing scenario,we first design a task delay optimization scheme that integrates computation offloading and federated learning for the edge machine learning task.After that,we focus on the edge collaborative computing scenario,and propose a coded subtask assignment scheme to optimize the task delay based on the consideration of the end-to-end dynamic characteristic.Finally,also for the edge collaborative computing scenario,we focus on the cache-enabled computing task,and reduce its task delay with the optimal service route selection and coding design.Specifically,the three major research parts and main contributions of this dissertation can be summarized as follows:1)Delay optimization for the edge machine learning task based on the multi-user collaboration.Due to the global collaborative mechanism in the edge machine learning task,the delay performances of the edge computing scheme and the federated learning scheme are limited by huge communication overhead and the computing straggler,respectively.To realize the task delay optimization,we propose an edge-user collaborative computing scheme,which combines computation offloading with federated learning and obeys the global collaborative computing mechanism.In this scheme,the computing stragglers can flexibly offload partial computation to the edge node to reduce the computational burden,thereby balancing the computing delay among users.After analyzing the task delay,we take the offloading data size as the optimization variable and formulate a power-constrained delay optimization problem.Based on the solution,a threshold-based offloading strategy is proposed.Moreover,we consider the situation where users may be offline during the computation.Based on the idea of user grouping,the offloading strategy is extended to a more general dynamic-user scenario.2)Delay optimization for the edge collaborative computing task with the consideration of the end-to-end dynamic characteristic.Due to the parallel computing structure,the task delay performance of the edge collaborative computing is limited by the straggling edge node.Due to the difficulty in predicting dynamic straggling nodes,it is challenging to develop the subtask assignment strategy for the task delay optimization.To this end,we propose a coded subtask assignment strategy,which avoids collecting the computation results of the straggling nodes by adding appropriate computational redundancy.We first analyze the dynamic factors of task arrival,queuing,transmission and computation stages,based on which an end-to-end stochastic system model is constructed.Then,according to the processing strategies in the edge node buffer,we further divide the coded subtask assignment strategy into non-purging and purging coded schemes.Although the expected task delay for both schemes are intractable,we obtain closed-form expressions for their respective lower and upper bounds.Finally,theoretical results show that the coded subtask assignment scheme has significant performance gain in task delay compared with the traditional uncoded scheme.3)Delay optimization for the cache-enabled edge collaborative computing task.For the cache-enabled computing task,data acquisition and task computation are coupled,which makes it difficult to improve task delay performance by optimizing cache or computing policy alone.To this end,we consider proposing a joint cache placement and computing offloading strategy.Since each strategy has its corresponding service route,the joint cache placement and computing offloading problem can be transformed into the service route selection problem.Specifically,we first apply the coded strategy to the edge caching and edge computing stages,which effectively mitigate the straggler effect.Then,according to the task processing flow of each service path,the corresponding task delay and resource overhead are analyzed.After that,we take the service path selection and coding design as the optimization variables,and formulate a task delay minimization problem with constrained caching,computing and communication resources.Finally,for the homogeneous and heterogeneous task scenario,we obtain the service path selection strategy and the problem solving algorithm that minimize the task delay.