IO-Gov: I/O 부하 반응형 CPU 거버너의 성능 및 탄소 배출량 최적화

초록

With increasing energy consumption of data centers driven by advancements in artificial intelligence technologies such as Large Language Models (LLMs), reducing power consumption and carbon emissions has become a critical challenge. Existing Linux CPU governors adjust frequencies based on CPU load. However, they do not take I/O-related metrics into account, which can lead to inefficient frequency scaling when a task has long I/O wait time. In this study, we proposed IO-Gov, a novel CPU governor that could dynamically adjust CPU frequency by analyzing I/O wait time (iowait) and I/O throughput variation (ΔBps) in real time. IO-Gov aims to maintain I/O performance while reducing carbon emissions. To evaluate its effectiveness, we conducted comparative experiments with existing Linux CPU governors across various I/O workloads. The results demonstrate that IO-Gov shows improvement in both performance and operational carbon emission reduction compared to existing Linux CPU governors for certain workloads in I/O-intensive environments.

키워드