Multi-source DNN task offloading strategy based on in-network computing

Lizi Hu, Yuhao Chai, Qin Li, Weiyuan Li, Yong Zhang

Abstract: As applications grow in scale, the centralized computing approach leads to excessive bandwidth requirements and high computational latencies. The traditional computing model regards the network as a transmission pipeline and has not fully explored the potential of network devices. At present, in-network computing is a new type of computing model that delegates application-layer processing functions to the network data plane. It can process traffic during transmission, reduce the cost of network bandwidth transmission, and alleviate the computing pressure and energy consumption of the cloud-side system. However, it needs to consider the redeployment cost. This paper proposes a multi-source DNN task offloading strategy based on in-network computing, which combines edge calculation, in-network computing and cloud computing. At the same time, it makes full use of the traditional routing nodes with no computing ability in the network. Particle swarm optimization (PSO) is used to solve the problem in the offloading scheme optimization. Service level agreement violation (SLAV) is introduced, and the network resources are offloaded in balance while the quality of service of users is satisfied. Simulated experiment results show that the proposed algorithm can reduce the cost and achieve convergence compared with the traditional offloading algorithm. In particular, we can find that there is an optimal deployment scheme in the network, which can make full use of the computing resources and bandwidth resources of the network, significantly reduce the computing pressure and transmission overhead of the whole network, and realize the balanced offloading of node resources.