A Dynamic FPGA Reconfiguration for Accelerating Machine Learning Framework with Image Service in OpenStack

Abstract

This paper discusses the feasibility of a dynamic FPGA reconfiguration by deploying FPGA images at run-time, with a particular focus on the offloading kernels to accelerate the training or inference phase of machine learning framework. It is useful to create FPGA images in advance and share them through repository because it takes a long time to build and is large in size. In this work, The Intel Arria10 GX FPGA card is used as a FPGA device, the TensorFlow as a machine learning framework and Glance, image service project in OpenStack, as a repository for storing and managing FPGA images. The OpenCL SDK tool chain presented by Intel corporation is used to implement FPGA images and client APIs for KeyStone and Glance are used to retrieve FPGA images from Glance. The result shows that our implementation can retrieve FPGA images, deploy them to the FPGA device and execute code that is offloaded to the FPGA device correctly.