FLEA - FIT-Aware Heuristic for Application Allocation in Many-Cores based on Q-Learning

This paper introduces FLEA, a novel allocation technique for many-core systems that uses Q-learning to improve system reliability based on Failure In Time (FIT) monitoring. FLEA considers task energy consumption and thermal behavior between neighbor processing elements (PEs) to determine the most suitable task allocation that minimizes PE FIT. Through a design phase that learns by trials, FLEA creates the Q-table. FLEA allocates and migrates tasks at runtime by consulting the Q-table, bypassing the need for executing complex heuristics. Compared to state-of-the-art allocation techniques, the results show that FLEA, in addition to increasing the MTTF, the mapping and migration heuristics reduce the thermal amplitude, peak temperature, and spatial thermal distribution.