Speaker
Description
In 2022, IT4Innovations was forced to temporarily turn off operated HPC systems because electricity prices increased enormously. Since then, the price has stayed high, and to avoid a potential next outage due to budget shortage, we investigated possible improvements to the Karolina system's energy efficiency. The behavior of accommodated CPUs and GPUs when executing workloads of a wide range of arithmetic intensities and instruction sets was analyzed in various hardware configurations to identify high power savings settings with limited impact on the performance of users' applications.
The talk will present the methodology used to identify the hardware configurations, power savings gain, and statistical evaluation of the impact of reduced resources on users' jobs.