WG Avalon public 10 juin 2024

Sébastien Valat (Inria AirSea) a donné le séminaire suivant.

Titre : Profiling mémoire, présentation de MALT et NUMAPROF

Résumé:
La mémoire est souvent devenue un problème majeur dans les applications.

Ceci pour la question des performances d’accès, mais aussi par les 
volumes à gérer au sein d’applications souvent de plus en plus 
dynamiques, complexes et développées à plusieurs sur des dizaines d’années.

Se posent alors les questions suivantes :

  – Comment se rendre compte de ses erreurs et de patterns 
problématiques facilement corrigibles ?
   – Comment trouver où elle est consommée lorsque j’atteins les limites 
de ma machine (malloc, variables globales, TLS) ?
   – Le NUMA, je fais comment pour savoir si je me suis trompé et où ?

Après ma thèse sur la gestion mémoire en contexte HPC (malloc, kernel, 
NUMA, multi-threading….) sur large simulation numérique j’ai eu 
l’opportunité de développer deux profileurs mémoires, Malt (MALloc 
Tracker) et Numaprof. Avec ces outils, j’ai tenté de reporter ce que 
j’ai pu comprendre le long de mon chemin et rendre visible ce que 
j’avais du mal à visualiser à l’époque dans les codes cibles inconnus 
avec lesquels j’interagissais. Outils maintenant open-sources pour le C 
/ C++ / Fortran (rust).

Je présenterais donc ces deux outils avec en principe quelques exemples 
d’observations obtenues.

Site web : https://memtt.github.io/

Présentation de l’orateur:
Après un parcours en physique des particules, j’ai bifurqué pour 
répondre à mes questions sur l’informatique en science avec une thèse au 
CEA en gestion de la mémoire des supercalculateurs et depuis cheminé au 
CERN et pour partie dans l’industrie du HPC sur les IOs. Je suis en ce 
moment à l’INRIA dans le domaine de la simulation océanique.
https://svalat.github.io/

WG Avalon 3 mai 2022

Élise Jeanneau a présenté l’exposé suivant.

Titre: SkyData, un nouveau paradigme pour la gestion de données

Résumé

Les systèmes de gestion de données traditionnels sont centrées sur les 
applications, plutôt que sur les données. Le projet SkyData propose de 
renverser la donne en gérant des données autonomes, capables de décider 
d’elles-mêmes leurs migrations et réplications. Le système qui en 
résulte est distribué, dynamique et fondamentalement différent des 
systèmes de gestion de données existants. Ce nouveau paradigme permet 
aux utilisateurs d’intégrer leurs données au système sans avoir à en 
céder le contrôle à un gestionnaire de données tier.

SkyData est un projet soumis à l’ANR. Cette présentation a pour but 
d’introduire les bases de la structure de SkyData, et de discuter 
certains usages possibles de ces données autonomes.

Pierre Jacquot: Evaluating unikernels for HPC applications

Pierre Jacqot
June 29th 2020, 14:30–15:30

Title: Evaluating unikernels for HPC applications

Abstract: Unikernels are lightweight single-application operating systems. They are designed to run as virtual machines, but some are able to run on bare metal too. They are quite popular in the system research community due to the increase of performance they can provide. By reducing the system call overhead and the OS-noise, they might be a good alternative to containers for HPC applications. This report evaluate the suitability of unikernels for HPC applications. This is done by conducting stability and performance studies with the Bots benchmarks and the Rodinias benchmarks. They are performed on multi-core architectures, on single node.

Arthur Chevalier: Optimization of software license placement in the Cloud for economical and efficient deployment

Arthur Chevalier
November 17th 2020, 14:30–15:30

Title: Optimization of software license placement in the Cloud for economical and efficient deployment

Abstract: Today, the use of software is generally regulated by licenses, whether they are free, paid for and with or without access to their sources. The world of licensing is very vast and poorly understood. Often we only know the version most widely used by the general public (a software purchase is equal to a license). The reality is much more complex, especially for large publishers. In this presentation I will present the impact and importance of managing these licenses when using software in a cloud architecture. I will show a case study to demonstrate the impact of dynamic license management and the need to propose new ways to manage and optimize software assets.

Titre: Optimisation du placement des licences logicielles dans le Cloud pour un déploiement économique et efficient

Résumé: Aujourd’hui, l’utilisation des logiciels est généralement réglementée par des licences, qu’elles soient gratuites, payantes et avec ou sans accès à leurs sources. L’univers des licences est très vaste et mal connu. Souvent on ne connaît que la version la plus répandue au grand public (un achat de logiciel est égale à une licence). La réalité est bien plus complexe surtout chez les grands éditeurs. Dans cette présentation je présenterai l’impact et l’importance de la gestion de ces licences lors de l’utilisation de logiciels dans une architecture Cloud. Je montrerai un cas d’étude pour prouver l’impact de la gestion dynamique des licences et la nécessité de proposer de nouvelles façons de gérer et optimiser un patrimoine logiciel.

WG Avalon : Analysis of energy consumption in a precision beekeeping system

Hugo Hadjur (joint work with Laurent Lefevre and Doreid Ammar (Aivancity group))
October 13, 2020 – 2.30pm – 3.30pm

https://dl.acm.org/doi/10.1145/3410992.3411010
Abstract:
Honey bees have been domesticated by humans for several thousand years and mainly provide honey and pollination, which is fundamental for plant reproduction. Nowadays, the work of beekeepers is constrained by external factors that stress their production (parasites and pesticides among others). Taking care of large numbers of beehives is time-consuming, so integrating sensors to track their status can drastically simplify the work of beekeepers. Precision beekeeping complements beekeepers’ work thanks to the Internet of Things (IoT) technology. If used correctly, data can help to make the right diagnosis for honey bees colony, increase honey production and decrease bee mortality. Providing enough energy for on-hive and in-hive sensors is a challenge. Some solutions rely on energy harvesting, others target usage of large batteries. Either way, it is mandatory to analyze the energy usage of embedded equipment in order to design an energy efficient and autonomous bee monitoring system. This paper relies on a fully autonomous IoT framework that collects environmental and image data of a beehive. It consists of a data collecting node (environmental data sensors, camera, Raspberry Pi and Arduino) and a solar energy supplying node. Supported services are analyzed task by task from an energy profiling and efficiency standpoint, in order to identify the highly pressured areas of the framework. This first step will guide our goal of designing a sustainable precision beekeeping system, both technically and energy-wise.

WG – Felipe de Souza: A Throughput Model for Data Stream Processing on Fog Computing

2019-07-09

Title: A Throughput Model for Data Stream Processing on Fog Computing

Speaker: Felipe Rodrigo de Souza (LIP, AVALON team)

Location: LIP, Meeting room M7 3rd floor

Schedule: 14:00

Abstract:

Today’s society faces an unprecedented deluge of data that requires processing and analysis. Data Stream Processing (DSP) applications are often employed to extract valuable information in a timely manner as they can handle data as it is generated. The typical approach for deploying these applications explores the Cloud computing paradigm, which has limitations when data sources are geographically distributed, hence introducing high latency and achieving low processing throughput. To address these problems, current work attempts to take the computation closer to the edges of the Internet, exploring Fog computing. The effective adoption of this approach is achieved with proper throughput modeling that accounts for characteristics of the DSP application and Fog infrastructure, including the location of devices, processing and bandwidth requirements of the application, as well as selectivity and parallelism level of operators. In this work, we propose a throughput model for DSP applications embracing these characteristics. Results show that the model estimates the application throughput with less than 1% error.

WG – Carlos Cardonha: Online scheduling in magnetic tapes

2019-07-02

Title: Online scheduling in magnetic tapes

Speaker: Carlos Cardonha (IBM Research Brazil)

Location: LIP, meeting room M7 3rd floor

Schedule: 14:30

Abstract: Magnetic tapes have been playing a key role as means for storage of digital data for decades, and their unsurpassed cost-effectiveness still makes them the technology of choice in several industries, such as media and entertainment. Tapes are mostly used for cold storage nowadays, and therefore the study of scheduling algorithms for read requests tailored for these devices has been largely neglected in the literature. In this article, we investigate the Linear Tape Scheduling Problem (LTSP), in which read requests associated with files stored on a single-tracked magnetic tape should be scheduled in a way that the sum of all response times are minimized. LTSP has many similarities with classical combinatorial optimization problems such as the Traveling Repairmen Problem and the Dial-a-Ride Problem restricted to the real line; nevertheless, significant differences on structural properties and strict time-limit constraints of real-world scenarios make LTSP challenging and interesting on its own. In this work, we investigate several properties and algorithms for LTSP and some of its extensions. The results allowed for the identification of 3-approximation algorithms for LTSP and efficient exact algorithms for some of its special cases. We also show that LTSPR, the version of the problem with heterogeneous release times for requests, is NP-complete. OLTSP, the online extension of LTSPR, does not admit c-competitive algorithms for any constant factor c, but we nevertheless introduce an algorithm for the problem and show through extensive computational experiments on synthetic and real-world datasets that different embodiments of the proposed strategy are computationally efficient and over-perform by orders of magnitude an algorithm being currently used by real-world tape file systems.

Mini-bio: Carlos Cardonha is a Research Staff Member at the Natural Resources Optimization group at IBM Research Brazil, with a Ph.D. in Mathematics (T.U. Berlin) and with a Bachelor’s and a Master’s degree in Computer Science (Universidade de São Paulo). His primary research interests are mathematical programming and theoretical computer science, with focus on the application of techniques in mixed integer linear programming, combinatorial optimization, and algorithms design and analysis to real-world and/or operations research problems.

WG – Cyril Seguin: SeeDep: Deploying Reproducible Application Topologies on Cloud Platform

2019-03-26

Title: SeeDep: Deploying Reproducible Application Topologies on Cloud Platform

Speaker: Cyril Seguin (LIP, AVALON team)

Location: LIP, Salle du conseil du LIP 3rd floor

Schedule: 14:30

Abstract:

As part of the scientific method, any researcher should be able to reproduce the experimentation in order to not only verify the result but also evaluate and compare this experimentation with other approaches. The need of a standard tool allowing researchers to easily generate, share and reproduce experiments set-up arises. In this talk, I’ll present SeeDep, a framework aiming at being such a standard tool. By associating a generation key to a network experiment set-up, SeeDep allows for reproducing network experiments independently from the used infrastructure.

WG – Houmani Zeina: Study and design of data-driven services/microservices discovery mechanisms

2019-01-29

Title: Study and design of data-driven services/microservices discovery mechanisms

Speaker: Houmani Zeina (LIP, Avalon team)

Location: LIP, council room 394 nord 3rd floor

Schedule: 14:30

Abstract:

— English version

Usual microservice discovery mechanisms are normally based on user needs (Goal-based Approaches). However, in today’s evolving architectures, several new microservices can be created. This makes the classic approach insufficient to discover the available microservices. That’s why customers need to discover the features they can benefit from before searching the available microservices in their domain. We will present a data-driven microservice architecture that allows customers to discover, from specific objects, the functionalities that can be exerted on these objects as well as all the microservices dedicated to them. This architecture, based on the main components of classic microservice architectures, adopts a particular communication strategy between clients and registers to achieve the desired objective.

— French version

Les mécanismes de découverte de microservices classiques sont normalement basés sur les besoins des utilisateurs (Goal-based Approches). Cependant, dans les architectures actuelles qui évoluent fréquemment, plusieurs nouveaux microservices peuvent être créés. Cela rend l’approche classique seule insuffisante pour découvrir les microservices disponibles. C’est pourquoi, les clients ont besoin de découvrir les fonctionnalités dont ils peuvent bénéficier avant de rechercher dans leur domaine les microservices disponibles. Nous allons présenter une architecture microservices pilotée par les données qui permet aux clients de découvrir, à partir d’objets spécifiques, les fonctionnalités qui peuvent être exercées sur ces objets ainsi que l’ensemble des microservices qui leur sont dédiés. Cette architecture, basée sur les composants principaux des architectures microservices classiques, adopte une stratégie de communication particulière entre les clients et les registres permettant d’atteindre l’objectif recherché.

WG – Mathieu Stoffel: Improving power-efficiency through fine-grain monitoring in HPC clusters

2019-01-22

Title: Improving power-efficiency through fine-grain monitoring in HPC clusters

Speaker: Mathieu Stoffel (LIG, CORSE team)

Location: LIP, meeting room M7 3rd floor

Schedule: 14:30

Abstract:

Nowadays, power and energy consumption are of paramount importance. Further, reaching the Exascale target will not be possible in the short term without major breakthroughs in software and hardware technologies to meet power consumption constraints.
In this context, this papers discusses the design and implementation of a system-wide tool to monitor, analyze and control power/energy consumption in HPC clusters.
We developed a lightweight tool that relies on a fine-grain sampling of two CPU performance metrics: instructions throughput (IPC) and last level cache bandwidth.
Thanks to the information provided by these metrics about hardware resources’ activity, and using DVFS to control power/performance, we show that it is possible to achieve up to 16% energy savings at the cost of less than 3% performance degradation on real HPC applications.