PACS: Private and Adaptive Computational Sprinting

Abstract

Computational sprinting is a class of mechanisms that enables a chip to temporarily exceed its thermal limits to enhance performance. Previous approaches model the optimization of computational-sprinting strategies as a mean-field game and calculate the game's static equilibrium strategies using dynamic programming. This approach has three main limitations: (i) a requirement for a priori knowledge of all system parameters; (ii) inflexibility in equilibrium strategies, necessitating recalculation for any system parameter change; and (iii) the need for users to disclose precise characteristics of their applications without data privacy guarantees. To address these issues, we propose PACS, a private and adaptive mechanism that enables users to independently optimize their sprinting strategies. Our experiments in a simulated environment demonstrate that PACS achieves adaptability, ensures data privacy, and provides comparable performance to state-of-the-art methods. Specifically, PACS outperforms existing approaches for certain applications while incurring at most a 10% performance degradation for others, all without having prior knowledge of system parameters.