ADARE: Adaptive Resource Provisioning in Multi-FPGA Edge Environments
Ian Kersz, Henry Piceni, Michael Guilherme Jordan, José Rodrigo Azambuja, Fernanda Lima Kastensmidt, Antonio Carlos Schneider Beck · 2024
This paper introduces ADARE, a novel resource provisioning framework optimized for multi-FPGA environments. Unlike traditional edge scenarios, which may not fully account for the multitenancy of heterogeneous and dynamic demands, ADARE integrates a fully adaptive resource management and load-balancing system. This framework selects the most suitable FPGA configurations for varying task demands and adjusts resource allocation dynamically to optimize energy consumption and makespan. Configurable to a range of operational parameters, ADARE excels across various computational scenarios by effectively managing the balance between workload types and FPGA capabilities. Demonstrations in a heterogeneous workload environment show ADARE outperforming traditional homogeneous approaches. On average, ADARE achieves a 2.0x improvement in makespan and a 2.4x reduction in energy consumption. These results highlight its adaptability and efficiency in managing complex multi-tenant edge environments.