A Study Of Power Control For HPC Systems Based On The POMPP Management Framework

Nowadays,Moore's Law is getting closer to its end as the semiconductor industry slows down its pace in advancing photolithography.Specifically,as new technology nodes continue to be introduced,there are fewer and fewer atoms in each transistor,and this physical limitation constrains the further development of semiconductor processes.Among the various negative effects that follow,the limitation on power consumption is one of the most critical.It is an inevitable wall in the pursuit of higher performance for current and future computer systems.The dramatic increase in power consumption of HPC systems has constrained their development and application to a certain extent,so it is essential to optimize the power consumption of applications while keeping the power consumption of HPC systems within a given budget.In order to utilize an HPC system more efficiently on a given power budget or within a given deadline,it is desirable to be able to accurately estimate the performance or power consumption of an application before optimizing it for operation on any given system.To simplify this process,this thesis presents a straightforward yet effective approach to help users of HPC systems clarify the relationship between performance and power consumption of their applications.It helps users manage power consumption and optimize performance on power-constrained HPC systems through the use of a power management framework and a concise and convenient DSL.This approach makes the process of analyzing,modeling,and managing the performance and power consumption of applications much easier.With the help of this frameworks,only a few trial runs are required to establish a very accurate model for the HPC system.This greatly improves efficiency and reduces the difficulty of using HPC systems on a limited power budget.