Container-FPGA based edge computing acceleration scheme with adaptive resource scheduling

Abstract

Nowadays, the computational intensity of mobile applications has increased so that the applications are often to be offloaded to an edge cloud for better user experience. Even if the application is offloaded to the edge cloud and can exploit better computing power, using only the CPU may not be enough to deliver the service in real-time. This has led to more attempts to utilize hardware accelerators (e.g., GPU, FPGA) for edge clouds. Especially, an FPGA has an advantage in processing streaming data due to its pipeline benefits, so that it is drawing attention as an accelerator in the edge. However, the characteristics of edge clouds, FPGAs, and Docker-Kubernetes environments should be considered when applying FPGAs to a Docker-Kubernetes based edge cloud. The most important problem is that a specific setup is needed to utilize the FPGA inside the Docker container. This setting should vary depending on the environment of each node, so applying the FPGA to a Docker-Kubernetes cluster, which can have nodes in a variety configures, requires a software module that can provide the setting adaptively to the environment of each node. Besides, it is important to utilize an FPGA more efficiently in edge clouds, that are not as resource-rich as the central cloud. In this perspective, researches on virtualized FPGA (vFPGA) that shares an FPGA across multiple applications has been conducted, but research on resource allocation schemes which takes into account the performance degradation caused by vFPGA's internal interference is insufficient. In addition, in an edge computing environment where many service providers are likely to deploy a variety of services, providing a standardized interface for communication with vFPGA is essential. However, the existing vFPGA studies do not provide the standard interface so that it cannot be directly applied to the edge cloud. In this thesis, we proposed and implemented an edge computing structure, including a software module that enables an FPGA as an application accelerator in a Docker-Kubernetes environment. The structure includes a module that enables applications inside the Docker container to be accelerated through the FPGA, a module that informs the type and status of the FPGAs in each node, and a vFPGA system that provides OpenCL as a communication interface. A resource scheduling scheme that takes into account vFPGA's internal interference for more efficient use of the resource in edge computing environments was also proposed. Throughout, we showed that in the proposed edge computing system, the service latency can be reduced by more than 50% compared to existing resource scheduling schemes by using the resource scheduling scheme.