Building the Infrastructure for Deploying FPGAs in The Cloud
Published in Springer: Hardware Accelerators in Data Centers, 2019
At the University of Toronto, we are building the infrastructure required to support the deployment of FPGAs at a large scale in a data centre. We consider FPGAs to be peers to the CPU-based servers rather than using them as accelerator slaves, which is the more common view. The goal is to enable the allocation and use of the FPGAs as computing resources in the same way that current servers are provisioned in a data centre. Our approach is to build on the existing knowledge and experience with provisioning software-based processing elements in the data centre and adapting the current infrastructure to the differences that FPGAs bring. This will incur minimal disruption and adjustment to how systems are currently being deployed. To support FPGAs in a data centre requires many layers of infrastructure to be developed and adapted. These layers range from the very lowest levels of infrastructure provided in the FPGA itself to the upper layers of resource provisioning and Conference Paperiguration in the cloud. In this paper, we describe our work towards building the various layers to support deploying FPGAs in the data centre.