Research On Incremental Computing Offloading For Edge Intelligence

Novel smart environments are driving increasing interest in deploying neural network models at edge nodes.Since neural network computing tasks tend to be computationally intensive,deploying neural network models on resource-constrained edge nodes is a huge challenge.The existing mainstream research method is to split the neural network model and upload them to the edge server on demand to perform neural network computing tasks collaboratively.However,this method of relying only on a single edge server is easily affected by unstable factors such as network jitter,and is extremely unstable.Moreover,when multiple neural network computing tasks are simultaneously offloaded to the same edge node,they may be affected by accidental factors such as concurrency conflict exceptions.In addition,when performing neural network computing tasks collaboratively,another problem that tends to arise is whether the collaborative computing methods are resilient to sudden failures of edge nodes.If not,the failure of an edge node will cause the failure of the neural network partition deployed on the edge node.To solve the above problems,firstly,this paper proposes Conflict-resilient Incremental Computation Offloading for Multiple Edge Nodes(CICO)to improve the execution efficiency of neural network computing tasks in edge intelligence environments.CICO can divide the neural network model into several partitions by layers,dynamically select cooperative targets in a cluster of trusted edge nodes,and cooperate with multiple edge nodes to perform neural network computing tasks.In addition,this paper designs an advanced locking mechanism for CICO to deal with problems such as concurrency conflict exceptions,which improves the robustness of neural network applications while improving the execution efficiency.Secondly,this paper proposes a Failure-resilient Prediction for Edge Intelligence(FPEI)method to improve the failure resilience of distributed prediction.This method does not require additional model reconstruction or retraining,and can improve the failure resilience of distributed predictions.As part of FPEI,this paper designs a failure detection mechanism to detect and locate failures.In addition,an adaptive replica management and prediction algorithm is designed,which can adaptively adjust the number of replica partitions according to the availability of each neural network partition to optimize redundancy.Finally,this paper designs and implements Incremental Computation Offloading and Prediction Framework for Edge Intelligence(ICOPF)in a real edge intelligence environment composed of edge devices with different hardware configurations.Support multi-edge node incremental computing offload and fault resilience prediction.The experimental results based on the ICOPF framework show that,compared with other state-of-the-art baseline methods,CICO can significantly improve the execution efficiency of the neural network model and significantly improve the robustness while guaranteeing the performance.Moreover,the experimental results also show that FPEI can provide failure resilience for CICO in various edge node failure scenarios while ensuring the performance and execution efficiency of CICO.