Performance analysis of adaptive resource allocation scheme for openCL-based FPGA virtualization system

Abstract

With the recent advanced performance of FPGA accelerators in many applications, the demand to use FPGA as a hardware accelerator is increasing. In response to the demand for FPGA, cloud providers offering FPGA-as-a-Service (FPGA) are increasing rapidly. FaaS users reduce purchase and develop- ment costs of FPGA accelerators by borrowing instance with virtualized computing resources, including FPGA and FPGA development tools. With the growth of FaaS, the demand for FPGA virtualization has also increased. The recent studies on FPGA virtualization usually assume plenty of FPGA resources and assign dedicated FPGA board to a user. However, this way of resource allocation limits the sharing of FPGA among users, resulting in inefficiencies of resource utilization and degradation of system per- formance. To avoid the waste of resources, partial reconfiguration (PR) based FPGA virtualization that supports the sharing of FPGA has been proposed. However, PR-based FPGA virtualization is having difficulty to expand the technology due to the lack of standardized interface. Requirements for efficient use of FPGA resources and provision of the standard interface have become more important, as cloud environments have recently shifted from resource-intensive and centralized form to limited-resource and distributed form, like Mobile Edge(Accessible) Cloud. Because OpenCL has become an industry standard for use of FPGA, this thesis proposes and imple- ments FPGA virtualization (vFPGA) system that supports the execution of OpenCL kernel. Moreover, we propose a run-time management system and dynamic resource allocation scheme to efficiently allo- cate virtualized FPGA resources in vFPGA environments. The proposed dynamic resource allocation scheme, which is called as Reconfiguration Overhead aware Resource Allocation (RORA), minimizes the latency of all services in the system, taking into account the reconfiguration overhead and fitting problem of vFPGA resources. The experimental results showed that the proposed vFPGA system accommodate multiple users at a lower cost than the system with non-virtualized FPGA. Moreover, the proposed dy- namic resource allocation scheme showed the reduction of service latency when it is compared to static allocation.