Agenda 1. 2. 3. 4. 5. Construction and Management High Avaibility How to use Software Batchsystem Compute-Cluster - 2 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 1. Construction and Management/1 Where to find: Grimm-Zentrum, Server Room, water cooled cabinet Aufbau: - 8 Supermicro Twin2-Server, 2 U 8 Supermicro Twin2-Server every has 4 nodes with 2 Intel-Nehalem CPUs every has 4 Cores, Infiniband QDR, 48 GB memory at all 32 nodes, 256 Cores, 1.536 TB memory Compute-Cluster - 3 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 1. Construction and Management/2 - two master nodes are responsible for login to the address: clou.cms.hu-berlin.de - they manage different services (DHCP, NFS, LDAP, Mail) - they start und stop alle nodes if necessary, only they communicate into the Universtiy network - they provide development software (Compiler, MPI) - and also application software - they organize the batch service Compute-Cluster - 4 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 1. Construction and Management/3 - two as a HA cluster configured server provide the parallel file system (Lustre FS) for temporary data of large size lulu:/work - all nodes are equipped with fast infiniband network (QDR) which is responsible for node communication during parallel computations - the master and file server nodes are monitored by central Nagios of CMS - the compute server are monitored local with Ganglia Compute-Cluster - 5 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 1. Aufbau und Cluster-Verwaltung/4 -/work Data - 1 MDT , 2 OSTs Stellt /work bereit - 3 OSTs - Login clou.cms.hu-berlin.de - Failover Login - 36 Port Infiniband Switch - node1 bis node4 - node5 bis node8 - node25 bis node28 - node29 bis node32 - /home Data Compute-Cluster - 6 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 2. High Avaibility /1 - all servers are equipped with redundant power supplies - master server: RedHat Resource Group Manger rgmanager configures and monitors the services on both master nodes - in case of failover of one master the other one takes over the services of the failed one - file server: for the Lustre FS /work are also redundant configured, if one fails the other one takes the MDS (Meta Data Server) and also the OSTs (Object Storage Targets), possible becuse all data are stored on a virtual SAN disk Compute-Cluster - 7 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 3. How to use/1 - all colleagues who have not enough resources at the institut are allowed to use the Kkk system - necessyary is an account alle at CMS, the acoount will be opend for the service - login on master clou.cms.hu-berlin.de with SSH form University network - you may login to nodes via ssh from master node without further authentification - Data storage: - /home/<institut>/<account> unique user home dir - /afs/.cms.hu-berlin.de/user/<account> OpenAFS home dir - /work/<account> working dir - /perm/<account> permanent dir - /scratch/<account> on nodes lokal working dir - Migration of data from old cluster via /home Verzeichnis or with scp Compute-Cluster - 8 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 3. How to use/2 Data saving: Kkk - data data data data data in in in in in /home alle daily into TSM, max. 3 older versions /afs/.cms.hu-berlin.de daily, the whole semester /perm daily on disk, max. 3 older versions /work no saving /scratch no saving - /work and /scratch are controlled on achieve high water mark of 85%, data older than 1 month willl be removed - important data should be copied to /perm or to a home dir - a parallel SSH is installed, for calling commands on all nodes pssh --help pssh –P –h /usr/localshare/nodes/all w |grep load Compute-Cluster - 9 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 4. Software/1 System software Kkk - Operating system: CentOS 5 = RedHat EL 5 alle Development Software - GNU Compiler - Intel Compiler Suite mit MKL - Portland Group Compiler - OpenMPI - MPICH Application software - Chemie: Gaussian, Turbomole, ORCA - Matlab, Maple - further Software is possible Compute-Cluster - 10 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 4. Software/2 Kkk versions must be loaded with the module command - all Software module available alle -> all available modules (Software) - Development: module load intel-cluster-322 -> loads Intel Compiler+ module load openmpi-14-intel -> loads OpenMPI ... - Application software: module load g09-a02 -> loads Gaussian09 A02 module load orca27 -> loads ORCA 2.7 module load matlab-10a -> loads Matlab R2010a ... Compute-Cluster - 11 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 5. Batch Service/1 Open Source Product: SUN Grid Engine (SGE) - cell clous is installedalle - on one master node the Qmaster daemon is running the other one is working as a slave and will get activ if the first fails - SGE supports parallel environments (PE) - There is a Grafical User Interface QMON - With QMON all configurations may be shown and actions may be done - The are two parallel environments installed: ompi (128 Slots) and smp (64 Slots) - These will be allocated to the parallel queues Compute-Cluster - 12 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 5. Batch Service/2 - parallelKkk jobs upto 32 cores are allowed, max. 4 Jobs a 32 Cores at the same time in bigpar alle - in queue par there are upto 8 jobs a 8 cores allowed, all jobs are running on a separate node - ser and long are serial queues - inter is for interactive computations (Matlab, GaussView) - all values and configurations are preliminary and may be changed on user requirements - batch scripts of old cluster can’t be used for the new one, because of new batch system Compute-Cluster - 13 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 5. Batch Service/3 Queue Kkk Priority Prozessors alle 8-32 Memory Slots PE Runtime 40 GB 128 ompi 48 h bigpar +5 par 0 4-8 20 GB 64 smp 24 h long -5 1 4 GB 32 - 96 h short +10 1 4 GB 32 - 6h inter +15 1 1 GB 8 - 3h Compute-Cluster - 14 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 5. Batch Service/4 Kkk - Users are collected to lists = working groups e.d.: cms, limberg, alle haerdle, ... - submition of jobs: example scripts for all applications are in /perm/skripte, they include the module calls - e.d. Gaussian09 computation: cd /work/$USER/g09 cp /perm/skripte/g09/run_g09_smp_8 qsub run_g09_smp_8 qstat qmon& -> Grafical User Interface Compute-Cluster - 15 -> ev. Änderungen Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 5. Batch Service/5 Kkk developing and starting (computation of Pi) - MPI program cd /work/$USER/ompi alle module load intel-cluster-322 module load openmpi-14-intel cp /perm/skripte/ompi/cpi.c . mpicc –o cpi cpi.c cp /perm/skripte/ompi/run_ompi_32 . qsub run_ompi_32 Enthält den Aufruf des MPI Programmes mpirun –np 8 –machinefile nodefile cpi cat nodefile -> enhält den Knoten, z.B. node8.local slots=8 Compute-Cluster - 16 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli 5. Batch Service/6 Kkkcommands - important qstat -> Status der eigenen Jobs alle qstat –u \* -> Liste aller Jobs im Cluster qdel <jobid> -> Entfernen des Jobs qconf –sql -> Liste aller Queues qconf –sq par -> Konfiguration Queue par qconf –sul -> zeigt die Userlisten (Gruppen) qconf –su cms -> User der Liste/Gruppe cms qconf –spl -> Liste der Parallel Environments qacct <jobid> -> Abrechnungsinfo zum Job qstat –q bigpar –f -> zeigt, ob Queue disabled Compute-Cluster - 17 Jahreskolloquium 29.06.2010 Daniela-Maria Pusinelli