Run-Time Monitoring and ML-Based Modeling in Reconfigurable Multi-Accelerator Systems
Juan Encinas, Alfonso Rodríguez, A. Otero, Eduardo de la Torre · 2021
Reconfigurable multi-accelerator systems usually have to deal with dynamically changing operating conditions when working in edge computing scenarios. To fully exploit execution performance while at the same time complying with the available power budget and application requirements, these systems need to deploy the optimal number and type of hardware accelerators at any point in time. In this paper, Machine Learning techniques are applied to extract predictive models of the execution of multiple hardware accelerators. These models can be later used to make decisions on the most appropriate system configuration to solve a given problem under specific timing and power constraints. Besides, a non-intrusive integrated instrumentation tool to measure power consumption and execution performance in FPGA-based systems is also proposed. This monitoring infrastructure is used to acquire the training data that feed the ML-based models. The proposed approach has been validated using the MachSuite benchmarks running on several multi-accelerator systems generated with ARTICo3, a custom high-performance embedded computing framework, and the Vitis commercial solution by Xilinx.