Optimisation of System Throughput Exploiting Tasks Heterogeneity on Space Shared FPGAs

Abstract

There are challenges in optimising system throughput in FPGA-based cloud computing due to mapping constraints resulting in suboptimal space sharing of resources, as the number of tasks grow and become more heterogeneous. This work proposes a methodology for exploring and optimising their resource utilisation. By identifying high-level synthesis parameters for each task, machine learning models and intelligent clustering are then employed to define clusters of tasks which will share the FPGA space. Assuming heterogeneity characterisation of tasks and thus static partitioning of the FPGA, it is ensured that each task in a cluster accommodates other tasks' resource requirements resulting in a higher compute density. Using 11 high performance computing tasks, we achieve an average 3.3× higher system throughput at 2.8× better energy efficiency when compared to existing approaches.