Dynamic Hardware Accelerator Selection Achieving Optimal Utilization of Resources

Edge computing with offloading over a network is being considered for enabling users to use advanced applications such as autonomous control applications. When network transmission time and/or task-processing time change, the total offloading time also varies. To guarantee low latency of offloading service in this case, over-performance hardware accelerator (HWA) must be assigned on edge computing, which leads to inefficient HWA resource allocation. To address this issue, we propose a method of selecting a HWA corresponding to change of the network transmission time and/or task-processing time. The proposed method refers past transmission times and input data size for the offload, enabling the prediction of changes in network transmission time and/or task-processing time. We evaluated the time for offloading through simulation and found that our method achieves a high percentage of offloads requested by users while efficiently using HWAs.