Calendar

Asymptotic Studies of Large Distributed systems with Failures

The qualitative behavior of a large scale storage network of non reliable file servers is investigated. In such systems a fraction of the processing capacity can be used to  duplicate  files on new servers when necessary.  Due to random losses, with probability 1 all files will be lost eventually if no new files are added. The main question in this context is of understanding the relation between the durability of the system and the parameters of the network: the duplication capacity, the failure rate, the duplication factor and the average load of a node. Different scalings via mean-field limits or stochastic averaging can be used to study the eventual decay of these networks. They will be reviewed and discussed. An analysis of placement algorithms will be also presented.

Detecting Performance Outliers for Task-based HPC Applications in
multi-[CPU|GPU|Node] clusters

Programming paradigms in High-Performance Computing have
been shifting towards task-based models which are capable of
adapting readily to heterogeneous and scalable
supercomputers. Detecting performance outliers in such environments
is particularly difficult because it must consider architecture
heterogeneity and variability. In this work we present how we have
employed a very simple performance model to highlight task outliers
of the well-known tiled-based dense Cholesky factorization running
on top of StarPU-MPI, a runtime for task-based applications. Such
work has been integrated into our visualization framework based on the
R programming language and the tidyverse meta-package. Experiments
have been conducted in a controlled environment using the Chifflet
cluster at Lille, part of the Grid'5000 infrastructure, using up to
eight nodes, each one equipped with 28 cores and two GPUs. The
preliminary results, derived from collected traces, indicate that
explicit binding for the MPI and GPU-managing threads, within
StarPU, alleviate the issue, leading to performance gains.

Les algorithmes de non-regret sont souvent utilisés dans les jeux répétés où les joueurs ont peu d’information sur le jeu auquel ils jouent. Ces algorithmes garantissent que le regret de chaque joueur est sous-linéaire. La moyenne temporelle des stratégies choisies en suivant un algorithme de non-regret converge dans l’ensemble des équilibres corrélés. Cependant cela ne donne aucune information sur la convergence de la séquence de stratégies.
Nous nous sommes intéressés à la question « est-ce que la sequence de stratégie obtenue pas un algorithme de non regret converge vers un équilibre de Nash? ».
Dans cet exposé, je présenterai un algorithme de non regret appelé Hedge qui est une version d’algorithmes à poids exponentiels. En particulier, je discuterai la convergence des séquences de stratégies obtenues par Hedge en utilisant deux types d’informations accessibles aux joueurs.