Comments on ”Improving the computing efficiency of HPC systems using a combination of proactive and preventive checkpoint”

In this short note, we provide some comments on the recent paper “Improving the computing efficiency of HPC systems using a combination of proactive and preventive checkpointing” by Bouguerra et al., published in [3]. We start by identifying some errors in their equations. Then we explain that they do not actually use the distribution of lead times, contrary to statements by the authors. Finally, we show that their algorithm does not change policy at the best possible moment, and we point to our own work [2] for the (correct version of the) optimal algorithm. Key-words: fault tolerance, checkpointing, prediction, algorithms, model, exascale ∗ LIP, École Normale Supérieure de Lyon, France † University of Tennessee Knoxville, USA ‡ Institut Universitaire de France § INRIA Commentaires sur l’article “Improving the computing efficiency of HPC systems using a combination of proactive and preventive checkpoint” Résumé : Dans cette courte note nous commentons l’article “Improving the computing efficiency of HPC systems using a combination of proactive and preventive checkpointing” de Bouguerra et al. [3]. Nous commençons par identifier des erreurs dans la mise en équation du problème. Nous expliquons ensuite que, contrairement à ce qu’ils prétendent, les auteurs n’utilisent pas la distribution du délai de prédiction (lead time). Finalement, nous montrons que leur algorithme ne change pas de politique au moment optimum, et nous indiquons que nous avons présenté l’algorithme optimal dans un rapport de recherche [2]. Mots-clés : Tolérance aux pannes, checkpoint, prédiction, algorithmes, modèle, exascale Comments on “Improving the computing efficiency using proactive and preventive checkpoint” 3