I/O-Cool: Using I/Os for Temperature Regulation When Training on the Edge

Abstract

Processing data directly at edge nodes reduces latency and enhances data privacy and security, but the lack of advanced cooling on these devices makes them unsuitable for sustained AI training, leading to overheating and accelerated hardware degradation. This paper is motivated by the observation that edge data intensive applications, in particular AI training, often alternate between I/O and compute phases. Notably, we found that processor temperatures consistently decrease during the I/O phases, revealing opportunities to optimize thermal behavior. We propose I/O-Cool, a proactive thermal management strategy for I/O–compute alternating edge workloads. I/O-Cool employs a lightweight predictive model combined with a Pareto-based selection strategy to determine the optimal data granularity to efficiently alternate I/Os and compute. This allows the system to maximize cooling opportunities through I/O phases while reducing the need for processor frequency reductions during the compute process. Experimental results demonstrate that I/O-Cool respects the imposed temperature limit, with a satisfaction rate of 90%-100%, that is up to 45% to 60% better than past work while maintaining competitive execution times.