Research On Service Placement And Task Offloading Of Model Training For Edge Intelligence

While we are into the 5G era,the large amounts of data generated at the edge and emerging delay-sensitive applications make it difficult for Cloud Computing paradigm to cope with.Edge Computing is proposed to meet these challenges.In order to make full utilization of the data generated at the edge,methods such as machine learning are usually adopted.Among these methods,deep learning has achieved satisfactory per-formance in many application scenarios represented by graphics processing.Under the concept of "Edge Intelligence",it becomes a popular technology trend to combine Edge Computing with such technologies as deep learning and conduct distributed models training at the edge,due to edge data's inherent geographically distributed feature.To carry out distributed model training at the edge,it is firstly required to place cor-responding model training services on edge nodes.While existing service placement studies often focus on light-weight common edge services,mainly from the perspec-tive of request delay and energy consumption.However,the task requests for model training service are distinguished from the ordinary ones of normal edge service due to their resource intensiveness and long-term operation time.Therefore,it is required to study service placement from a new perspective.This paper considers the placement of model training services from the perspective of time period,and studies the optimal service placement strategy based on the two quantitative methods of placement cost and placement revenue to maximize the system's overall revenue.The service placement problem is firstly formalized as a nonlinear 0-1 integer programming problem.By con-version of optimizing target,relaxation,delaying constraint judgment,and randomized rounding techniques,the RDSP algorithm is proposed,whose performance guarantee is confirmed through theoretical analyses.Simulation results based on multiple perfor-mance criterion show that,compared to the baseline algorithms,the RDSP algorithm can increase the overall system revenue by 26%while maintaining absolute advantages over others on load balancing performance.In order to ensure the overall performance,the edge system is also required to prop-erly offload distributed model training tasks.A typical edge distributed model training task often involves multiple data nodes,it's needed to tackle the data nodes assignment problem when offloading the task.However,in most existing related studies on task of-floading,there is only one request node for a task,and in most existing related studies on edge model training,they often assume that the data is always ready on edge nodes and ignore the assignment process of data nodes.In addition,due to the limited resources on edge nodes,it is necessary to study how to allocate resources for multiple distributed model training tasks at multiple edge nodes to ensure the quality of service,which is not considered by most existing work.This.paper jointly considers the offloading and resource allocation problem for edge distributed model training tasks.From the per-spective of maximizing the throughput of system training tasks,this joint problem is formalized as a nonlinear mixed integer programming problem that is proved to be NP-hard.Based on techniques such as constraint transformation and randomized rounding,the RDJOA algorithm is designed.Theoretical analyses show that the RDJOA algo-rithm's performance is close to optimal.Furthermore,simulation results indicate that,compared to the baseline algorithms,the RDJOA algorithm can improve the system throughput by 56%and resource utilization by 53%.