Invited Paper: On the Granularity of Bandwidth Regulation in FPGA-Based Heterogeneous Systems on Chip

Main memory sharing in commercial, FPGA-based Heterogeneous System on Chips (HeSoCs) can cause significant interference, and ultimately severe slowdown of the executing workload, which bars the adoption of such systems in the context of time-critical applications. Bandwidth regulation approaches based on monitoring and throttling are widely adopted also in commercial hardware to improve the system quality of service (QoS), and previous work has shown that the finer the granularity of the mechanism, the more effective the QoS control. Different mechanisms, however, might exploit more or less effectively the available residual memory bandwidth, provided that the QoS requirement is satisfied. In this paper we present an exhaustive experimental evaluation of how three bandwidth regulation mechanisms with coarse, fine and ultra-fine granularity compare in terms of exploitation of the system memory bandwidth. Our results show that a very fine-grained regulation mechanism might experience worse system-level memory bandwidth exploitation compared to a coarser-grained approach.