ETCF – Energy-Aware CPU Thread Throttling and Workload Balancing Framework for CPU-FPGA Collaborative Environments

Abstract

Warehouses and Cloud Servers have been adopting collaborative CPU-FPGA architectures as an alternative to enable extra acceleration for data-parallel applications by distributing the workload among both devices. However, exploiting the benefits of this environment is challenging, since the amount of data distributed to each device affects the needed CPU processing power and, therefore, the number of active CPU threads for the task. In this scenario, activating the highest number of CPU threads, which is usually the choice of programmers, will not always achieve the best possible performance or energy consumption. To address this challenge, we propose ETCF – Energy-Aware CPU Thread Throttling and Workload Balancing Framework for CPU-FPGA collaborative environments. ETCF automatically provides efficient CPU-FPGA execution by selecting the right workload balance and the number of CPU threads for a given collaborative application. ETCF framework offers different optimization goals: performance, energy, or energy-delay product (EDP). Compared to the baseline (an equally balanced workload executing with the maximum number of CPU threads), ETCF provides, on average, 93% of EDP reduction. We also show that ETCF achieves near-optimal solutions by comparing it to an Oracle, but just taking 3.32% of its searching time.