Dynamic Computation Offloading Strategy with DNN Partitioning in D2D Multi-Hop Networks
Xin Guo, Chongwu Dong, Wushao Wen · 2021
The expansion of smart mobile applications has posed great challenges on mobile devices with limited computation resources. Since DNN based applications are usually computation-intensive, it is hard for resource-poor mobile devices to meet delay requirements. Inspired by the DNN model partition strategy, the paradigm of computation offloading in multi-hop D2D networks allows a device to complete tasks with the assistance of other devices. However, with the increase of the number of devices, the complexity of network architecture brings great difficulties to the designment of computation offloading strategy. To cope with this situation, we propose one D2D task offloading strategy based on the complex network theory to find the optimal task offloading assignment. Our proposed strategy can mitigate the routing congestion when transmitting tasks in D2D multi-hop networks, meanwhile it can dynamically find the partition point of DNN models to minimize the overall delay. We finally conduct extensive evaluations and demonstrate the effectiveness of our strategy.