WP3 Étude et modélisation de l’ordonnancement et du déploiement des applications 7/5/2009
-
Upload
brynne-stanley -
Category
Documents
-
view
22 -
download
0
description
Transcript of WP3 Étude et modélisation de l’ordonnancement et du déploiement des applications 7/5/2009
WP3 Étude et modélisation de l’ordonnancement et du déploiement des
applications
7/5/2009
WP3: Overview
• Les tâches T3.1: Etude et modélisation T3.2: Mise en œuvre
• Bilan
2
Ordonnancement
E. Caron - Réunion #11 - 7/5/09
Ocean-Atmosphere scheduling within DIET
4
• Improve performances in a climate prediction application• Modelization of the application• Proof of usage of Grid’5000 and DIET
Scheduling on real application • Scheduling done at two levels
Groups of processors at cluster level Distribution of scenarios at grid level
• Real implementation suffered from technical limitations• Simulations are quite precise but we need to keep one resource
for post-processing tasks
E. Caron - Ocean-Atmosphere scheduling within DIET - APDCT-08
Cluster Level Scheduling – Experimental result
• Experiment: 10 scenarios, 5 clusters, from 11 to 112 resources• Every resource is taken into account• Average makespan is strictly decreasing when adding more
resources• The decrease rate of the average makespan diminishes
E. Caron - Ocean-Atmosphere scheduling within DIET - APDCT-08
5
Result Grid Level Scheduling
6
• Comparison with Round Robin on 5 clusters• Maximum speedup: 25%• With a higher load, the
algorithm behaves better with a few resources
• Convergence on gains• Gain of 25% ≈ 230h on
a ≈ 822h long experiment
E. Caron - Ocean-Atmosphere scheduling within DIET - APDCT-08
Workflow Management
• Workflow representation Direct Acyclic Graph (DAG)
Each vertex is a task Each directed edge represents communication between
tasks
• Goals Build and execute workflows Use different heuristics to solve scheduling
problems Extensibility to address multi-workflows
submission and large grid platform Manage heterogeneity and variability of
environment
• Research topics addressed Workflow scheduling with parallel tasks Multiple workflows scheduling (makespan
minimization and fairness optimization)
• Specific agent for workflow management (MA DAG)• Two modes:
MA DAG defines a complete scheduling of the workflow (ordering and mapping) MA DAG defines only an ordering for the workflow execution, the mapping is done in the
next step by the client which pass by the Master Agent to find the server where execute the workflow services.
• Design of heuristics for mixed parallelism
Architecture with MA DAG
agent
SeD_parallel
Front-end
NFS
OAR LSF PBS Loadleveler
GLUE
SeD_batch
SeD_seq
Parallel and batch submissions
• Parallel & sequential jobs → transparent for the user
• Submit a parallel job→ system dependent
NFS: copy the code? Numerous batch systems Batch schedulers behaviour
(queues, scripts, etc.) Information about the
internal scheduling process Monitoring
& Performance prediction
Simulation (Simbatch)
agent
SGE
Task reallocation in a grid environment
• Batch simulator: Simbatch [Y. Caniou, J.S. Gay] Validated against OAR (less than 2% error) Based on Simgrid
• Grid simulated by Simgrid with several Simbatch instances• Different algorithms studied: MCT MinMin MaxMin on batchs
using FCFS or CBF - MaxMin and MinMin can not try to reschedule more than 30 jobs at each reallocation (choose oldest or youngest jobs)
• Comparison of jobs completion time with and without reallocation
Task reallocation in a grid environment
• Traces of one month from Grid’5000 (january to june 2008) on three sites
• Reallocation triggered every hour
Batch 1 Batch 2
Meta - SchedulerMeta - Scheduler
Get waiting jobs
Submit Cancel
Reschedule
Déploiement
E. Caron - Réunion #11 - 7/5/09
Deployment Brick: ADAGE
• Automatic deployment tool for grid environment• Only one command to deploy
3 kinds of input information Resource description application description control parameter
• Planning model (random, round-robin), …
• Plug-in for generic application mapping RR, Random, DIET, Graal-heuristics
• Plug-in for each application kind Description convector Configuration of application CCM, MPI, JXTA, P2P, DIET, GFARM, SEQ
Plug-in: from 400 to 4700 C++ lines
META Enable constraints between any other application kinds (at the generic
level)
13
Identification of the steps of Automatic Deployment
14
MPI Application Description
CCM Application Description
Resource Description
Generic Application Description
Control Parameters
Deployment Planning
Deployment Plan Execution
Application Configuration
Stat
ic
Applic
atio
ns
Deployme
nt Tool
Comparaison GoDIET / Adage
• Déploiement sur Grid’5000 : Entre 25 et 305 nœuds Entre 1 et 8 grappes Heuristique pour créer automatiquement la hiérarchie DIET
Comparaison GoDIET / Adage
• Hiérarchie DIET générée
17
ADAGE & LEGO
• Clean implementation of the Adage model UML-like based specifications Separations of planner and application plugins from core
• Extension of the internal generic model (GADe) Support of graph-like generic description
In particular recursive structures like trees (for DIET)
• Support of pseudo-dynamic re-deployment• Support of the G5K API• Working and stable tool
Use to deploy CCM, JuxMem & DIET elements Cf Demonstrator talk
Grid'5000 Reservation Utility for Deployment Grid'5000 Reservation Utility for Deployment UsageUsage
• Web: http://grudu.gforge.inria.fr
GRUDU – Resources Allocation
• We are able to reserve ressources (OAR1 & OAR2) Time parameters, date and reservation walltime Queue OARGrid sub behaviour/ Script to launch
GRUDU – Monitoring
• We are able to monitor the status of the grid/site/a job.• We are able to get instantaneous/historical data with Ganglia
GRUDU - KaDeploy/JFTP
• GUI for KaDeploy jobs deployment• File Transfert interface (local<->remote/rsync on Grid'5000)
WP3: Bilan
• Réalisations principales Délivrable D3.2
ADAGE DIET 2.3
GRUDU• Perspectives
Prise en compte automatique de la plateforme pour le planning Clustering auto-stabilisant
ADAGE ? Utilisation de l’ordonnancement de l’application du CERFACS pour un
modèle régional atmosphérique: CRIP UJF (IMAG. Grenoble) D’autres classes d’applications à ordonnancer
Ordonnancement et gestion de données: création et utilisation de DAGDA
• Collaborations Salomé (EDF) [thèse en cours] Université de Picardie Jules Verne Université du Nevada Las Vegas Université d’Hawaii
22