Build power profiling tool for modern CPUs

Abstract EN

Reduce the application power consumption is one of the main challenges for the HPC community.

Code power profilers are very important for researchers to identify the performance bottlenecks and power consumption for their code.

Most of the modern CPUs are equipped with a built-in sensor to allow researchers and HPC engineers to estimate the power consumption of the running applications.

To estimate the power consumption for any piece of code running on CPU, you need to eliminate the confounding factors as possible and run the code many times until the average converge.

The reason for that is the environment, which has the OS and other processes and services running at the same time with your code and may report incorrect power readings.

In this paper, we build a power profiler tool, which saves the researcher time by running and profiling different pieces of code with different types of workloads, and keeps running until the average converge.

Furthermore, we identify and eliminate the environment confounding factor which saves the researcher time and gives a realistic result for power consumption experiments.