Title : A Reconfigurable Component Model for HPC

Speaker: Vincent Lanore

Abstract: High-performance applications whose structure changes dynamically during execution are extremely complex to develop, maintain and adapt to new hardware. Such applications would greatly benefit from easy reuse and separation of concerns which are typical advantages of component models. Unfortunately, no existing component model is both HPC-ready (in terms of scalability and overhead) and able to easily handle dynamic reconfiguration. We aim at addressing performance, scalability and programmability by separating locking and synchronization concerns from reconfiguration code. To this end, we propose directMOD, a component model which provides on one hand a flexible mechanism to lock subassemblies with a very small overhead and high scalability, and on the other hand a set of well-defined mechanisms to easily plug various independently-written reconfiguration components to lockable subassemblies. We evaluate both the model itself and a C++/MPI implementation called directL2C.

WG_2015-04-29_Vincent

Title : Design of Survivable VPN Topologies over a Server Provider Network

Speaker: Brigitte Jaumard

Abstract: In the context of multiple-hop working routing for IP layer traffic requests. The design problem is composed of two problems which are simultaneously solved: (i) Finding the most efficient or economical multi-hop routing of the IP traffic flows with different bandwidth granularities over the logical topology, which involves some traffic grooming, (ii) Ensuring that the logical topology is survivable throughout an appropriate mapping of the logical links over the physical topology, if such a mapping exists. In order to solve such a complex multi layer resilient network design problem, we propose a column generation ILP model. It allows exploiting the natural decomposition of the problem and helps devising a scalable solution scheme. We conducted numerical experiments on a German network with 50 nodes and 88 physical links. Not only we could solve much larger data instances than those published in the literature, but also observe than multi-hop routing allows a saving of up to 10% of the number of lightpaths, depending on the traffic load.

Title : Vers un modèle de composants supportant l’ordonnancement de tâches pour le calcul de haute performance

Speaker: Jérome Richard

Abstract: Les applications de haute performance ont une durée de vie souvent plus grande que celle des plate-formes sur lesquelles elles reposent. L’adaptation de ces applications à différentes plate-formes est un processus nécessaire, long et coûteux. Les composants logiciels offrent de nombreux avantages de génie logiciel simplifiant l’adaptation des applications. Parallèlement, on souhaiterait garder de bonnes performances à travers les adaptations. Les modèles d’ordonnancement de graphes de tâches permettent de tirer parti efficacement des architectures hétérogènes tout en apportant des performances portables. Cet présentation propose et évalue un modèle de composants avec ordonnancement de tâches visant à profiter des avantages des deux approches sur des SMP. Les résultats montrent que le modèle proposé dispose d’avantages provenant des approches à composants (séparation des préoccupations) et des approches à tâches (équilibrage de charge).

Title : Gysela5D, Adapting a GYrokinetic SEmi-LAgrangian code for current architectures and towards Exascale

Speaker: Julien Bigot

Abstract: In order to design and operates the future reactor for nuclear fusion such as ITER (tokamaks), physicists need to better understand the various types of instabilities that develop in the plasma and impact the confinement of heat. Simulation of Ion Temperature Gradient (ITG) instabilities based on the Vlasov equations require huge amounts of computational power with a discretization of both the spacial and velocity space (6D). The gyrokinetic approximation makes this kind of simulation possible by reducing this to “only” 5D. Up to now, the semi-Lagrangian code Gysela5D has been used to perform large simulations using a few thousands cores (8k to 16k cores typically). These simulations make the hypothesis that electrons are adiabatic but recent advances seem to indicate that some instabilities could only be explained by simulating kinetic electrons. In order to do that, the spacial mesh would have to be refined by a 60³ ratio and time steps by a 60 ratio. Such simulations would require Exascale capable machines. In this talk, I present some challenges identified in order to provide an Exascale-ready code as well as solutions recently implemented and work in progress to tackle these. I especially focus on three such piece of work:

• memory scalability optimization;
• I/O optimizations for both checkpoints and result writing;
• communication patterns optimization for big number of cores (Blue Gene/Q).

I will also present recent results that show that the code scales with good performance up to 1,835,008 threads (the complete Juqueen Blue Gene/Q at Jülich).

PDF: WG_150224_jbigot-gysela

Title: Reconstructable Software Appliances with Kameleon

Speaker: Salem Harrache

Abstract: A software appliance builder bundles together an application with its needed middleware and an operating system to allow easy deployment on Infrastructure as a Service (IaaS) providers. These builders have the potential to address a key need in the computer science community: the ability to reproduce an experiment. This talk presents a software appliance builder called Kameleon that automates the construction of complex software appliances targeted at research on operating systems, HPC and distributed computing, Devops etc. It does so by proposing a highly modular description format that encourages shareability and reuse of procedures. Moreover, it provides debugging mechanisms for improving experimenter’s productivity.

PDF: WG_150127_Salem