arXiv subjects

Luiz Angelo Barchet-Estefanel

Publications and source records attributed to Luiz Angelo Barchet-Estefanel.

4 recordsLinked to original sources

arXiv2005

Prédiction de Performances pour les Communications Collectives

Des travaux récents visent l'optimisation des opérations de communication collective dans les environnements de type grille de calcul. La solution la plus répandue est la séparation des communications internes et externes à chaque grappe, mais cela n'exclut pas le découpage des communications en plusieurs couches, pratique efficace démontrée par Karonis et al. [10]. Dans les deux cas, la prédiction des performances est un facteur essentiel, soit pour le réglage fin des paramètres de communication, soit pour le calcul de la distribution et de la hiérarchie des communications. Pour cela, il est très important d'avoir des modèles précis des communications collectives, lesquels seront utilisés pour prédire ces performances. Cet article décrit notre expérience sur la modélisation des opérations de communication collective. Nous présentons des modèles de communication pour différents patrons de communication collective comme un vers plusieurs, un vers plusieurs personnalisé et plusieurs vers plusieurs. Pour évaluer la précision des modèles, nous comparons les prédictions obtenues avec les résultats des expérimentations effectuées sur deux environnements réseaux différents, Fast Ethernet et Myrinet.

cs.DC↗

arXiv2004

Identifying Logical Homogeneous Clusters for Efficient Wide-area Communications

Recently, many works focus on the implementation of collective communication operations adapted to wide area computational systems, like computational Grids or global-computing. Due to the inherently heterogeneity of such environments, most works separate "clusters" in different hierarchy levels. to better model the communication. However, in our opinion, such works do not give enough attention to the delimitation of such clusters, as they normally use the locality or the IP subnet from the machines to delimit a cluster without verifying the "homogeneity" of such clusters. In this paper, we describe a strategy to gather network information from different local-area networks and to construct "logical homogeneous clusters", better suited to the performance modelling.

cs.DC↗

arXiv2004

Fast Tuning of Intra-Cluster Collective Communications

Recent works try to optimise collective communication in grid systems focusing mostly on the optimisation of communications among different clusters. We believe that intra-cluster collective communications should also be optimised, as a way to improve the overall efficiency and to allow the construction of multi-level collective operations. Indeed, inside homogeneous clusters, a simple optimisation approach rely on the comparison from different implementation strategies, through their communication models. In this paper we evaluate this approach, comparing different implementation strategies with their predicted performances. As a result, we are able to choose the communication strategy that better adapts to each network environment.

cs.DC↗

arXiv2004

Performance Characterisation of Intra-Cluster Collective Communications

Although recent works try to improve collective communication in grid systems by separating intra and inter-cluster communication, the optimisation of communications focus only on inter-cluster communications. We believe, instead, that the overall performance of the application may be improved if intra-cluster collective communications performance is known in advance. Hence, it is important to have an accurate model of the intra-cluster collective communications, which provides the necessary evidences to tune and to predict their performance correctly. In this paper we present our experience on modelling such communication strategies. We describe and compare different implementation strategies with their communication models, evaluating the models' accuracy and describing the practical challenges that can be found when modelling collective communications.

cs.DC↗