Uncore Frequency Impact on Performances and Energy Consumption
Rémi Dulong (Université de Neuchâtel)
Since the beginning of the history of computing, global computing capacity has grown exponentially over time. Today, we are able to run many really demanding applications, ranging from neural networks training to realistic simulations of chaotic systems.
However, constant improvement of our computers and supercomputers has a worrying side effect. Indeed, the power needed to perform all these computations has also never been as high as today. In a world where energy will become more and more precious and expensive, High Performance Computing must be energetically optimized as much as possible.
The energy consumption is now one of the multiple parameters that has to be considerate by hardware manufacturers. As the processor can be one of the most consuming component inside a computer, new architectures and features about energy monitoring appeared. Now, on most of recent processor architectures, the working frequency of each CPU core is independent, and all other components outside these CPU cores are associated to a new frequency, called “Uncore frequency”.
This talk is about our research on the impact of this Uncore frequency onto processors performances and energy consumption. We tried to show how this frequency can be modified in order to get a good trade-off between these two measures. Then, we managed to find if the type of application used could alter the impact of Uncore frequency changes.
We will show that for some applications, we managed to reduce the socket energy consumption by 16% with absolutely no overhead on the execution time of the application.
Rémi Dulong got a Master degree in computer science with a minor in embedded systems in Telecom SudParis (Evry, France).
He started his PhD in the University of Neuchâtel on October 2018. This work is the result of an internship in the SAMOVAR laboratory, with the HP2 team in Telecom SudParis, supervised by Amina Guermouche.
Back to EBSIS Events section.