Prediction Based Run-Time Reconfiguration on Many-Core Embedded Systems

This paper studies prediction based run-time system reconfiguration strategy to tolerate environment change and hardware malfunction on many-core embedded systems. System reconfiguration will invoke application migration, which may significantly impact system's timing behaviors, therefore, it is vital important to select an appropriate migration strategy after which the system's performance is still acceptable. The essence of our prediction based approach is to pre-estimate the impact of possible migration strategies and upon which to choose the optimal one. Our proposed approach includes data profiling, model training and execution time prediction phases. The initial data profiling and model training are conducted in the design stage, and will be continuously updated as time goes on after the system is in use. When system reconfiguration is invoked, the most recently trained models will be used for prediction at run-time. Extensive experiments have been set up by running multiple benchmarks on a four-core hardware platform and experimental results evaluate and validate our proposed approach.