WCET-Aware Partitioning and Allocation of Disaggregated Networks for Multicore Systems

The integration of machine learning into safety-critical cyber-physical systems has significantly increased computational demands, which are often met by modern multicore platforms. While complex memory subsystems, including local caches, make it challenging to maintain timing predictability, they also provide opportunities for worst-case execution time (WCET) optimization through improved data locality. To address this, we propose a multicore partitioning and allocation strategy that leverages sparse structures through neural network disaggregation to optimize the WCET. Our evaluation shows that disaggregated neural networks achieve a significantly reduced WCET, compared to fully connected monolithic neural networks of similar size.