Run your HPC jobs in Eco-Mode: revealing the potential of user-assisted power capping in supercomputing systems
Fuente:
arXiv
Salvato in:
| Autori principali: | , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2024
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866909160505868288 |
|---|---|
| author | Angelelli, Luc Carastan-Santos, Danilo Dutot, Pierre-François |
| author_facet | Angelelli, Luc Carastan-Santos, Danilo Dutot, Pierre-François |
| contents | The energy consumption of an exascale High-Performance Computing (HPC) supercomputer rivals that of tens of thousands of people in terms of electricity demand. Given the substantial energy footprint of exascale HPC systems and the increasing strain on power grids due to climate-related events, electricity providers are starting to impose power caps during critical periods to their users. In this context, it becomes crucial to implement strategies that manage the power consumption of supercomputers while simultaneously ensuring their uninterrupted operation.This paper investigates the proposition that HPC users can willingly sacrifice some processing performance to contribute to a global energy-saving initiative. With the objective of offering an efficient energy-saving strategy by involving users, we introduce a user-assisted supercomputer power-capping methodology. In this approach, users have the option to voluntarily permit their applications to operate in a power-capped mode, denoted as 'Eco-Mode', as necessary. Leveraging HPC simulations, along with energy traces and application metadata derived from a recent Top500 HPC supercomputer, we conducted an experimental campaign to quantify the effects of Eco-Mode on energy conservation and on user experience. Specifically, our study aimed to demonstrate that, with a sufficient number of users choosing Eco-Mode, the supercomputer maintains good performances within the specified power cap. Furthermore, we sought to determine the optimal conditions regarding the number of users embracing Eco-Mode and the magnitude of power capping required for applications (i.e., the intensity of Eco-Mode). Our findings indicate that decreasing the speed of jobs can decrease significantly the number of jobs that must be killed. Moreover, as the adoption of Eco-Mode increases among users, the likelihood of every job to be killed also decreases. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2404_03271 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Run your HPC jobs in Eco-Mode: revealing the potential of user-assisted power capping in supercomputing systems Angelelli, Luc Carastan-Santos, Danilo Dutot, Pierre-François Optimization and Control The energy consumption of an exascale High-Performance Computing (HPC) supercomputer rivals that of tens of thousands of people in terms of electricity demand. Given the substantial energy footprint of exascale HPC systems and the increasing strain on power grids due to climate-related events, electricity providers are starting to impose power caps during critical periods to their users. In this context, it becomes crucial to implement strategies that manage the power consumption of supercomputers while simultaneously ensuring their uninterrupted operation.This paper investigates the proposition that HPC users can willingly sacrifice some processing performance to contribute to a global energy-saving initiative. With the objective of offering an efficient energy-saving strategy by involving users, we introduce a user-assisted supercomputer power-capping methodology. In this approach, users have the option to voluntarily permit their applications to operate in a power-capped mode, denoted as 'Eco-Mode', as necessary. Leveraging HPC simulations, along with energy traces and application metadata derived from a recent Top500 HPC supercomputer, we conducted an experimental campaign to quantify the effects of Eco-Mode on energy conservation and on user experience. Specifically, our study aimed to demonstrate that, with a sufficient number of users choosing Eco-Mode, the supercomputer maintains good performances within the specified power cap. Furthermore, we sought to determine the optimal conditions regarding the number of users embracing Eco-Mode and the magnitude of power capping required for applications (i.e., the intensity of Eco-Mode). Our findings indicate that decreasing the speed of jobs can decrease significantly the number of jobs that must be killed. Moreover, as the adoption of Eco-Mode increases among users, the likelihood of every job to be killed also decreases. |
| title | Run your HPC jobs in Eco-Mode: revealing the potential of user-assisted power capping in supercomputing systems |
| topic | Optimization and Control |
| url | https://arxiv.org/abs/2404.03271 |