Proposal and Evaluation of GPU Offloading Parts Reconfiguration During Applications Operations for Environment Adaptation

In recent years, not only CPUs with few cores but also heterogeneous hardware such as GPUs, FPGAs, and multi-core CPUs are increasingly used in many applications. However, to fully utilize these, users need to have technical knowledge that covers hardware such as CUDA. To overcome this high technical barrier, we have proposed environment-adaptive software that enables high-performance operation by automatically converting application code written for normal CPUs by engineers in accordance with the deployed environment and by setting appropriate amounts of resources. So far, we have also verified the elemental technologies that automatically offload to GPU and FPGA before the start of operation. Until now, we only considered conversions and settings before the start of operation. In this paper, we verify that the logic is reconfigured in accordance with the usage characteristics during operation. Especially for GPU logic, there is no example of reconfiguration during operation, so the proposed method can be expected to have a great impact on clouds or similar businesses. We propose a GPU reconfiguration method during operation and find that the application running on the GPU is reconfigured to other offload loops or other offload applications in accordance with the current usage trends. Through a reconfiguration experiment, performance improvement and break time are measured, and the effectiveness of the method is demonstrated.