Improving Supercomputer Usage with Aging Awareness
Résumé
Lifetime of electronic devices has a critical impact on their environmental footprint. In addition, the high-demand by AI companies of GPU has reduced tremendously their availability for supercomputing centers. Consequently, improving the duration of CPUs and GPUs is becoming a major issue in High Performance Computing (HPC) domain. This paper investigates how one can obtain more FLOP from a single machine before a fatal failure and what the trade-offs with performance are. The lifetime of computing devices has been shown to be strongly connected with the temperature and thus with the running frequency of usage. We investigate how one could use node frequency reconfiguration to optimize the usage of HPC machines. We discuss several aging models to estimate when there is a benefit to have a dedicated scheduling algorithm, and when simply selecting a correct constant machine frequency is enough. We show that a correct decision can increase considerably the number of FLOP that one can get from a single machine with a trade-off in terms of average usage frequency. Because aging models are currently inaccurate, we consider different models and discuss the robustness of our algorithms to this inaccuracy. We show that they can still be used with gain. These results show the importance of research in having better aging models and also in taking into consideration aging in our usage of supercomputers.
Domaines
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |