Machine Learning-based Thermally-Safe Cache Contention Mitigation in Clustered Manycores
DescriptionThis paper introduces SmartCM, the first technique that mitigates cache contention in clustered manycores through application migration to maximize performance under a thermal constraint. To achieve this goal, SmartCM uses a neural network (NN) model to predict the impact of possible migrations on performance considering cache contention before performing the thermally-safe migration with the highest performance gain. Additionally, cluster-level dynamic voltage and frequency scaling (DVFS) is employed to exploit available thermal margins and to prevent any potential thermal violations. Compared to the state-of-the-art techniques, SmartCM achieves up to 47% performance gains with an average of 11%, while maintaining thermal safety.
TimeWednesday, July 12th2:25pm - 2:40pm PDT
Location3002, 3rd Floor
Timing and Low Power Design