FPGA Virtualisation on Heterogeneous Computing Systems --- Model, Tools, and Systems

Abstract

Field Programmable Gate Arrays (FPGAs) can be utilised to speed up applications by two orders of magnitude as compared to running on conventional CPUs. However, designing FPGA accelerators still remains challenging for most users. Furthermore, despite the trend of integrating FPGA resources into high-performance and cloud computing systems, FPGA management is still immature commonly following a run-tocompletion execution model and with providing no abstraction layers to underlying hardware. Thus, FPGA virtualisation is highly desired to provide an abstraction level for FPGA development and deployment in complex heterogeneous CPU+FPGA computing systems which provide us an opportunity to adapt the workload on either software (running on CPUs) or hardware (running on FPGAs). This PhD thesis aims at promoting FPGA virtualisation for such heterogeneous computing systems ranging from embedded, to edge, to high-performance and cloud nodes. Consequently, this thesis proposes a fully FPGA-virtualised computing model to tackle these obstacles on complex heterogeneous CPU+FPGA systems. Moreover, partial reconfiguration is one of the key techniques to implement the proposed model, yet has still significant limitations which prevented us to implement such model on real hardware. These limitations motivated this PhD project and resulted in the development and implementation of several solutions to overcome these limitations, and hence, advance the partial reconfiguration technique towards the propsed fully FPGA-virtualised computing model. Combining this new capability of partial reconfiguration with other academic design tools, a novel design methodology has been developed to realise the proposed model on real hardware. Finally, resulting systems of the proposed design methodology on various heterogeneous computing platforms have shown significant improvements in compute performance thanks to the here implemented FPGA-virtualised techniques.

Date of Award	1 Aug 2020
Original language	English
Awarding Institution	The University of Manchester
Supervisor	James Garside (Supervisor) & Dirk Koch (Supervisor)