PowerCoord: Power Capping Coordination for Multi-CPU/GPU Servers using Reinforcement Learning

Sustainable Computing: Informatics and Systems(2020)

引用 5|浏览24
暂无评分
摘要
•We identified multi-CPU/GPU servers create unique challenges for power capping.•First, meeting the cap must involve coordination among the various power domains.•Second, workload characteristics must be included in the budget for each domain.•Third, these servers host multiple jobs that must be taken into account.•Our proposed power cap controller, PowerCoord, considers all three challenges.•We propose multiple policies for different scenarios of workload characteristics.•We propose a reinforcement learning's algorithm to choose among policies.•Our controller shows effective operation on a real server with negligible overhead.
更多
查看译文
关键词
Power capping,GPGPU acceleration,Reinforcement learning
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要