PARIS*: Programming parallel and distributed systems for large scale numerical simulation applications Kerrighed, Vigne Christine Morin IRISA/INRIA * Common project with CNRS, ENS-Cachan, INRIA, INSA, Université de Rennes 1
Members of PARIS Project (sept 05) Scientific leader T. Priol (DR INRIA) Researchers F. André (Prof IFSIC) G. Antoniu (CR INRIA) J-P. Banâtre (Prof IFSIC) M. Bertier (MdC INSA) L. Bougé (Prof ENS) Y. Jégou (CR INRIA) A-M. Kermarrec (DR INRIA) C. Morin (DR INRIA) J.L. Pazat (MdC INSA) C. Perez (CR INRIA) Post-docs A. Viana A. Ribes G. Vallée Engineers D. Margery (IR INRIA) P. Morillon (IE IFSIC) Engineers P. Gallard (DGA) V. Lefèvre (G5K - IFSIC) R. Lottiaux (DGA) G. Mornet (G5K - PRIR) P. Palosaari (CoreGRID) J. Parpaillon (Ing. Associé) PhD candidates H-L. Bouziane (INRIA 2) J. Buisson (MENRT 3) Y. Busnel (ENS 1) L. Cudennec (INRIA-Région 1) M. Fertré (MENRT 1) M. Jan (MENRT 3) E. Jeanvoine (CIFRE 2) S. Lacour (INRIA 3) E. Le Merrer (CIFRE FT) S. Monnet (INRIA-Région 3) Y. Radenac (MENRT 3) E. Riviere (MENRT 2) L. Rilling (ENS 3)
Studied Systems Clusters A set of interconnected PC used as a single computing resource Grid A set of resources (processor, memory, disk, ) interconnected via Internet P2P systems A dynamic distributed system without any global state
Research Directions Single system image operating system Problem: clusters are difficult to program/use Challenge: to give the illusion that a cluster is a single machine Component based middleware Problème: code coupling applications are complex Challenge: How to facilitate the design of such applications while providing high performance? Advanced programming models Problème: Current programming models are not adequate for highly dynamic systems Challenge: How to express computing/coordination is such an environement? Data sharing service Problème: data sharing in large scale grids Challenge: sharing mutable data Systèmes P2P Problème: Master and optimize a P2P system Challenge: Characterize a P2P system and searching relevant information Experimental grid platform Problème: Need to experiment to validate our research results Challenge: building a reconfigurable grid
Grid 5000 Experimental Platform Contribution to the construction of Grid 5000 9 sites, 5000 processors Rennes site 500 processors (powerpc, Xeon, Opteron) Dual processor nodes Participants Researcher: Y. Jégou Engineers: V. Lefèvre, D. Margery, P. Morrillon, G. Mornet 500 500 1000 500 500 500 500 500 500 Grid-5000 Rennes
International Collaborations (funded) Cluster OS University of Ulm (Germany), ORNL (USA), Rutgers University (USA) Grids Pisa University (Italy), SNU (Korea) Large scale data management UIUC (USA) P2P systems Vrije Universiteit (The Netherlands)
OS for Clusters and Grids Kerrighed Single System Image (SSI) operating system for high performance computing on clusters Vigne Operating system to ease the use and programming of grids
Kerrighed Objectives Virtual shared memory multiprocessor Global and transparent resource management Tolerating node failures transparently for the applications High performance Approach Design of distributed OS mechanisms within an existing OS (Linux)
Kerrighed Achievements Customizable efficient full SSI operating system for high performance computing on clusters Small clusters (up to 256 nodes) Advanced research prototype Integration of the work of 3 Ph.D. students (R. Lottiaux (2001), Geoffroy Vallée (2004), P. Gallard (2004)) Robust prototype able to execute real applications provided by EDF R&D and DGA Open source software www.kerrighed.org Stable version (K V1.0.2) based on Linux 2.4.29 Demo LiveCD based on Knoppix kerrighed.users@irisa.fr integrated in OSCAR ssi-oscar.irisa.fr OSCAR is a snapshot of methods for building, programming, and using clusters. It consists of a fully integrated and easy to install software bundle designed for high performance cluster computing.
Efficient Operating System Comparison with other SSI for clusters OpenSSI, openmosix Results published in CC-GRID 2005 Internship of Benoît Boissinot Efficient communication system Highly reactive communication system to support Kerrighed distributed operating system services Compatibility of Kerrighed with efficient communication drivers used by HPC applications (such as GM for Myrinet)
Collaborations EDF R& D (since 2000) PhD and post-doc grants DGA (2003-2005) Funding for research engineers ORNL (S. Scott) Integration of Kerrighed in OSCAR University of Ulm (M. Schöttner) Fault tolerant SDSM University of Rutgers (L. Iftode) High availability Invited researchers R. Badrinath (IIT Kharagpur) Isaac Scherson (UCI)
Current Research Directions Fault tolerance Large scale parallel application checkpointing System initiated checkpoints Checkpointing grid applications Master & PhD Thesis of Matthieu Fertré High availability Current work of Pascal Gallard and Renaud Lottiaux Tolerating hot node addition and eviction Phenix Investigating the backdoor approach in the context of Kerrighed Master Thesis of Benoît Boissinot Application SSI cluster OS Node 1 Node 2 Node 3
Technology Transfer KerLabs (http://www.kerlabs.com) Start-up funded by Pascal Gallard and Renaud Lottiaux Software suite based on Kerrighed technologies EasyAdmin: Global cluster management EasyCheckpoint: Checkpoint/restart of parallel applications EasyRun: Application deployment & scheduling on clusters EasyCluster: the whole Kerrighed SSI solution Optimized support for high performance networking technologies Open source model
Vigne: a Grid OS Design and implementation of a Grid OS to ease the use and programming of very large grids Highly decentralized system Algorithms based on local knowledge Self-healing system Dealing with multiple quasi-simultaneous reconfigurations Single System Image Flexible system
Vigne Infrastructure based on decentralized overlays Structured and unstructured overlays Application manager for reliable application execution Resource discovery & allocation service On-going work (PhD Thesis of Emmanuel Jeanvoine) Volatile data sharing service PhD Thesis of Louis Rilling Complex application deployment PhD Thesis of Boris Daix (co-advised with Christian Pérez) To start beginning of 2006
Collaborations EDF R&D PhD grants
Future Work XtreemOS Project Integrated Project (FP6 - Call 5) Goal Under evaluation Building and Promoting a Linux-based Operating System to Support Virtual Organisations for Next Generation Grids 18 partners Academic & industrial partners 8 countries (including China)
XtreemOS Main Objectives We will design, implement, evaluate and distribute an open source Grid OS with native support for virtual organizations. Development of a Grid Operating System Enhance Linux to support VO across multiple administrative domains Manage very large and Self-organizing and selfhealing system Available on PC, SMP, clusters, PDA and mobile phones XtreemOS software: 3 flavours Standard flavour for PC Federation flavour based on Kerrighed SD flavour for small devices XtreemOS software will make the VO management easy for administrators and work, within VOs, easy, secure and efficient. Application Appli Appli Appli Middleware Experimentation and evaluation with a comprehensive set of real usecases provided by ISVs and endusers Linux Computer Linux Computer XtreemOS Linux Computer Linux Computer Integration in notorious Linux distributions Mandriva, Red Flag Linux Building a reference open source Grid OS
Talks Kerrighed Pascal Gallard Matthieu Fertré Vigne Emmanuel Jeanvoine JuxMem Sébastien Monnet