Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Experience using Slurm on ARIS HPC System
Nikos Nikoloutsakos
GRNETGreek Research and Technology Network, Greece
hpc.gnret.gr
27 September 2016
Slurm User Group Meeting 2016 27 September 2016 1/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Overview
1 IntroductionWho we areSystem Overview
2 SlurmConfigurationAdministration - Monitoring
3 Issues
4 Feature request
Slurm User Group Meeting 2016 27 September 2016 2/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Who we areGreek Research Technology NetworkGRNET enables researchers from Greece to obtain access to the powerfulnational High Performance Computing system ARIS.
Advanced Research Information SystemARIS Infrastructure provides state-of-the-art supercomputing capabilitiesfor large-scale scientific applications.
GRNET provides services to:Greece - Greek Academic Community
Greek UniversitiesTechnological institutionsResearch centers
EuropePRACE (Tier 1 system)DECIother EU Projects (Vi-seem, Eudat, EGI,...)
Slurm User Group Meeting 2016 27 September 2016 3/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Adiministrative and Application support
Support Team HPC provides:Management of InfrastructureUser Support
Comprehensive end-user supportUser support in operational problemsDocumentationEducational and Training Events
Application Support - Transfer and optimizing applicationPeer-Review support and coordination
Slurm User Group Meeting 2016 27 September 2016 4/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Open Access
Peer-Review AccessThe criteria for the evaluation:
Scientific ExcellenceImpact of the proposed researchThe need for HPC resourcesMaturity and experience of the principal investigator andhis/her teamFeasibility of the project based on a technical evaluationand the availability of resources
Slurm User Group Meeting 2016 27 September 2016 5/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Project TypesPreparatory-Development ProjectsExecution of scalability tests, performance tests, resolve issues.Code porting, development, optimization.
Review: Technical onlyCall: Always OpenAccess: 2-4 months
Production ProjectsProjects that have the technical expertise to take advantage ofavailable resources and are selected by the procedure of peerreview
Review: Technical-ScientificPeriodic Call - 2 per yearAccess: 1 year
Slurm User Group Meeting 2016 27 September 2016 6/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
since June 2015
First pilot operational phase in June 2015150 projects400 Users24 Organizations300 software modules120.000 jobs submitted,46M core hours (1 year)25 scientific publications (up to now)https://hpc.grnet.gr/results-publications/
Slurm User Group Meeting 2016 27 September 2016 7/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
ARIS
Compute Power: 180 TFlops (HPL) #465 Top500 - iterationJune 2015
Slurm User Group Meeting 2016 27 September 2016 8/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
ARIS - Compute Nodes I
426 compute nodes: IBM NextScale n360 M48520 cores: 2x (Intel E5 [email protected] - 10 core) pernode27TB total memory: 64GB memory per node(8 RDIMMS, 1866 MHz)Half-width, 1U systems grouped in 6U enclosures(12 nodes per enclosure)
Slurm User Group Meeting 2016 27 September 2016 9/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
ARIS - Compute Nodes II
6 Racks, 6 enclosures per rack.DisklessIBM 1PB GPFS, Tape Library ΙΒΜ TS3500 6PBMax nominal power consumption: 162 KW(154 KW on HPL). 183 KW with air-cooling.
Slurm User Group Meeting 2016 27 September 2016 10/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
ARIS - Network
Mellanox SX6536 648-Port Infiniband Director SwitchFDR 56 Gbits / secFat tree non-blocking mode450 QSFP+Optical cables5 Km fabric cables
Slurm User Group Meeting 2016 27 September 2016 11/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
ARIS - Expansion44 gpu nodes: “2 x NVIDIA Tesla k40m” acceleratednodes.
Dell Power Edge R7302 x Intel Xeon [email protected] GB RAM
18 phi nodes: “2 x INTEL Xeon Phi 7120p” acceleratednodes.
Dell Power Edge R7302 x Intel Xeon [email protected] GB RAM
44 fat nodesDell PowerEdge R8204x Intel Xeon [email protected] GB RAM
IBM 1PB GPFS,Tape Library ΙΒΜ TS3500 6PB
Slurm User Group Meeting 2016 27 September 2016 12/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
ARIS - Managment
14 support nodes, NextScale x3650 M4 2 x E5-2640v22x Managment Nodes, 2x Login Nodes, 10x service nodesMonitoring software xCAT, Nagios, Ganglia, BMS(Business Management System) Dell OpenManage, MRTGScheduler SLURM 14.11.8XDMoD, UMGMT (User Managment Tool) in house
Slurm User Group Meeting 2016 27 September 2016 13/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Partition Queues
ONE cluster ”ARIS”
Partition Description Nodescompute Thin nodes 426gpu GPU nodes 44phi PHI nodes 18fat FAT nodes 24taskp Serial queue 20
Default timelimit 2 days
Slurm User Group Meeting 2016 27 September 2016 14/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Configurtion parameters I
Consumable ResourcesSelectTypeParameters= CR_CORE_MEMORYShared mode - unless user specifies --exclusive
Resource LimitsAccountingStorageEnforce = associations,limits,safe
Generic Resource (GRES) SchedulingGresTypes = gpu,micmic offload mode only
Slurm User Group Meeting 2016 27 September 2016 15/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Configurtion parameters II
MpiDefault = pmi2Supports MPI implementation being used on system:Intelmpi,OpenMPI, mvapich2
The larger the job, the greater its job size priority.PriorityFavorSmall=NO
Accounting GatherAcctGatherEnergyType=acct_gather_energy/ipmiAcctGatherInfinibandType=acct_gather_infiniband/ofed
JobAcctGatherType = jobacct_gather/linux
Slurm User Group Meeting 2016 27 September 2016 16/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Priority Flags I
Multifactor PriorityPriorityType= priority/multifactorPriorityWeightAge = 100PriorityWeightFairShare = 1000PriorityWeightJobSize = 1000PriorityWeightPartition = 0PriorityDecayHalfLife = 00:00:00PriorityUsageResetPeriod = WEEKLYPriorityMaxAge = 30-00:00:00
PriorityWeightQOS=0
Slurm User Group Meeting 2016 27 September 2016 17/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Priority Flags II
Fair Tree FairsharePriorityFlags = FAIR_TREEPriorityCalcPeriod = 02:00:00
Backfill SchedulingSchedulerType= sched/backfill
Slurm User Group Meeting 2016 27 September 2016 18/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
$mybudget
$myreport
Slurm User Group Meeting 2016 27 September 2016 19/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
XdMod
Slurm User Group Meeting 2016 27 September 2016 20/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
UMGMT I
Users Management Tool
Tool to manage project proposals and user access on thesystem.Associate project proposals to slurm accountinginformationKeep Track start end dates per project,Extensions: core hours-access periodProject status , send alert emails to usersStatistics consumed core-hours(%) per project
in development: Ruby on Rails
Slurm User Group Meeting 2016 27 September 2016 21/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
UMGMT II
Slurm User Group Meeting 2016 27 September 2016 22/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
SLURM - MRTG
Slurm User Group Meeting 2016 27 September 2016 23/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Slurm Script TemplateHelps users prepare batch job scripts for Slurm at ARIS.
Acknowledgment BYU Job Script Generatorhttps://github.com/BYUHPC/BYUJobScriptGenerator
Slurm User Group Meeting 2016 27 September 2016 24/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Issues
Problem:Reservation (daily) had 20 nodes , 15 where active , 5where active by same user but for other job1 node (from 15) died, unable to reschedule.
Slurm User Group Meeting 2016 27 September 2016 25/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Feature request
More verbose error messages:Users could figure why a job is rejected.More information about which limit violatedMPI Task 0: may need more memoryAbility to specify less processes on first node.Allocation per GRES(gpu,mic) not only cpu ch
What’s Nextupgrade to version 16
Slurm User Group Meeting 2016 27 September 2016 26/27
SLUG 2016
N. Nikoloutsakos
IntroductionWho we areSystem Overview
SlurmConfigurationAdministration -Monitoring
Issues
Featurerequest
Thank you !
Slurm User Group Meeting 2016 27 September 2016 27/27