Cross-Layer Design Based Communication Optimization In Edge Networks

Recently,in order to effectively meet the massive and growing demand for cloud services,edge computing,i.e.,processing data in the vicinity of user equipments(UEs),has received increasing attention.Edge computing applications usually rely on the par-ticipation of a large number of UEs and generate high data transmission requirements.Therefore,as the core facility of edge computing,the access and communication ca-pabilities provided by the edge network will greatly affect the service quality of edge computing applications it carries.However,the current edge network still has some problems such as lack of access capability of some UEs,limited performance of large-scale concurrent communications,and insufficient support for intelligent computing.To this end,based on the idea of cross-layer design and collaboration between modules,this paper has conducted in-depth research on the above issues,and the re-search results obtained include the following aspects:(1)Considering it is difficult for IEEE 802.11 based UEs to access the edge net-work in harsh environments,we adopted a cross-layer design between the physical-layer and the MAC-layer(medium access control layer),and proposed a middleware solution Rateless802.11 based on adaptive error correction.Rateless802.11 can be eas-ily deployed on common commodity 802.11 devices.It concatenates LT codes with802.11 convolutional codes in the transmitter side to introduce proper redundancies in an incremental manner,so that the uncorrupted bits of MAC protocol data units can be adequately exploited,enhancing the error correction capability significantly.In re-ceiver side,it adopts a well-designed process flow of received packets,within which a belief-propagation based integrated decoder is employed,that decodes convolutional codes and LT codes together in a joint manner,leading to much less information-loss and decoding delay than serial decoding approaches.Experimental results show that Rateless802.11 can significantly improve the communication efficiency of 802.11 de-vices in harsh environments.Compared with state-of-the-art solutions,its throughput can be increased by several tens of times,which greatly enhances the access capabilities of these devices in the edge network.(2)Consider the concurrent communications in millimeter wave(mm Wave)based massive Machine Type Communications(m MTC),which is a typical scenario of edge network.We found that,due to the limited scattering effect of mm Wave as well as very dense UEs,existing multiuser detection(MUD)solutions based on sum-product algo-rithm or its variants running over factor-graphs,which is the key technique to provide concurrent communications,do not perform well.So we further proposed an approach named Bst Sum Prod,which is based on the collaboration between radio frequency(RF)and baseband related modules.According to the mm Wave channel status fed back by RF related modules,Bst Sum Prod first optimizes the factor-graph used for MUD based on node-split and node-contraction.In baseband related modules,it then takes a dynamic-programming based method to approximate the messages passing on the resulted factor-graph,which can improve the performance of sum-product algorithm significantly,and improves the throughput of concurrent communications.Extensive experimental results show that,compared to two state-of-the-art MUD approaches with polynomial-time complexities,Bst Sum Prod can achieve a higher throughput by 50%and 20%,respectively.(3)Considering the performance of federated learning in the edge network is lim-ited by its frequent aggregation process for client-side updates,we proposed the scheme Phy Arith based on cross-layer collaboration between the physical-layer and the application-layer.By exploiting the superimposed RF signal in the multiuser multiple-input multiple-output(MU-MIMO)enabled edge network,Phy Arith is taken to accelerate the aggre-gation process in federated learning,and improve its calculation performance.In Ph-y Arith,clients encode their local updates into aligned digital sequences which are con-verted into RF signals for sending to the server simultaneously,and the server directly recovers the exact summation of these updates as required from the superimposed RF signal by employing a customized sum-product algorithm.Phy Arith is compatible with commodity devices due to the use of full digital operation in both the client-side en-coding and the server-side decoding processes,and can also be integrated with other updates compression based acceleration techniques.Simulation results show that Ph-y Arith further improves the communication efficiency by 1.5 to 3 times for training Le Net-5,compared with solutions only applying updates compression.