- Infos im HLRS Wiki sind nicht rechtsverbindlich und ohne Gewähr -
- Information contained in the HLRS Wiki is not legally binding and HLRS is not responsible for any damages that might result from its use -

Powersched(Hawk): Difference between revisions

From HLRS Platforms
Jump to navigationJump to search
Line 3: Line 3:


==== Powersched ====
==== Powersched ====
Powersched proposes a reliable, efficient, and transparent mechanism to keep a HPC system under a power cap. The tool aims to increase the energy efficiency of available resources by enabling dynamic power capping based on an application-specific, energy-optimized runtime environment. Powersched can dynamically set a power cap for corresponding job nodes based on the application’s behavior at runtime. The power capping here is appied dynamically and will adapt to different application phases. Depending on the job characteristic we aim for an improvement in energy efficiency of order 20% with a moderate performance impact (or gain) of order 5%.  
Powersched proposes a reliable, efficient, and transparent mechanism to keep a HPC system under a power cap. The tool aims to increase the energy efficiency of available resources by enabling power capping based on an application-specific, energy-optimized runtime environment. The power capping is governed by a given metric (a variant of the Energy-Delay-Product) and will adapt to different application phases. Depending on the job characteristic we aim for an improvement in energy efficiency of order 20% with a moderate performance impact (or gain) of order 5%.  


A summary is included below.
A summary is included below.

Revision as of 14:38, 9 January 2024

Power Management (Powersched)

Powersched

Powersched proposes a reliable, efficient, and transparent mechanism to keep a HPC system under a power cap. The tool aims to increase the energy efficiency of available resources by enabling power capping based on an application-specific, energy-optimized runtime environment. The power capping is governed by a given metric (a variant of the Energy-Delay-Product) and will adapt to different application phases. Depending on the job characteristic we aim for an improvement in energy efficiency of order 20% with a moderate performance impact (or gain) of order 5%.

A summary is included below.

Overview and Architecture

Media:HLRS_Workshop_Powersched_GSS.pdf