Quickstart Guide

This is a quick start guide to help you start to work on the HPC-clusters CARL and EDDY.

If you have questions that arent answered in this guide, please contact the servicedesk of the IT services: servicedesk@uni-oldenburg.de.

HPC Cluster Overview

The HPC cluster, located at the Carl von Ossietzsky Universität Oldenburg, consists of two clusters named CARL and EDDY. They are connected via FDR Infiniband for parallel computations and parallel I/O. CARL uses a 8:1 blocking network topology and EDDY uses a fully non-blocking network topology. Further, they are connected via an ethernet network for management and IPMI. They also share an GPFS parallel file system with about 900TB net capacity and 17/12 GB/s paralell read/write performance. Additional storage is provided by the central NAS-system of the IT-services.

Both clusters are based on the Lenovo NeXtScale system.

CARL (271 TFlop/s theoretical peak performance):

• 327 compute nodes (9 of these with a GPU)
• 7.640 CPU cores
• 77 TB of RAM
• 360TB local storage

EDDY (201 TFlop/s theoretical peak performance):

• 244 compute nodes (3 of these with a GPU)
• 5.856 CPU cores
• 21 TB of RAM

For more detailed informations about the cluster, you can visit our Overview.

Login

If you want to access the HPC-cluster, you need to have an authorized university account. If you are not authorized yet, request an account.

You can use a SSH client of your choice or the command line on linux computers to connect to the cluster via ssh. To do so, use either

carl.hpc.uni-oldenburg.de

or

eddy.hpc.uni.oldenburg.de

For further informations about the login, please look at the guide located on the page Login to the HPC cluster.

File System

The cluster offers two files systems: The GPFS Storage Server (GSS) and the central storage system of the IT services.

GPFS Storage Server (GSS):

• parallel file system
• total (net) capacity is about 900TB
• R/W performance is up to 17/12 GB/s over FDR Infiniband
• can be mounted using SMB/NFS
• used as the primary storage for HPC (for data that is read/written by compute nodes)
• no backup!

Central storage system of the IT services (Isilon Storage System):

• NFS-mounted $HOME-directories
• high availability
• snapshots
• back
• used as permanent storage!

HOME:

• Path: /user/abcd1234
• Environment variable: $HOME

DATA:

• Path: /gss/data/abcd1234
• Environment variable: $DATA

WORK:

• Path: /gss/work/abcd1234
• Environment variable: $WORK

Further information can be found on the related page in the wiki: File system and Data Management

Software and Environment

There are many pre-installed software packages like compilers, libraries, pre- and postprocessing tools and further applications provided. We are using the module command to manage them.

With this command you can:

• list the available software
• access/load software (even in different versions)

Example: Show the software on CARL and EDDY and load the Intel compiler

[abcd1234@hpcl001 ~]$ module avail
-----------/cm/shared/uniol/modules/compiler-----------
... icc/2016.3.210
[abcd1234@hpcl001 ~]$ module load icc/2016.3.210
[abcd1234@hpcl001 ~]$ module list
Currenty loaded modules: ... icc/2016.3.210 ...

Basic Job Submission

The new workload manager and job management queueing system on CARL and EDDY is called SLURM. SLURM is a free and open-source job scheduler for Linux and Unix-like kernels and is used on about 60% of the world's supercomputers and computer clusters.

To submit a job on the HPC cluster you need two things:

• the command sbatch
• a jobscript

If you have your jobscript (an example is linked at the end) you can simply queue it with the command:

sbatch -p carl.p my_first_job.job

The option "-p" defines the used partition. Please keep in mind that choosing the right partition will allow your job to be run faster. If you choose the wrong one it might take a while for your job to start. Therefore we recommend you to look at wiki-article about partitions.

If you did submit your job sucessfully, you can check its status with

squeue -u abcd1234

Informations like JOBID, PARTITION, JOBNAME, USER, TIME and the amound of NODES will be displayed. (not adding the options -u will show every job that is currently running on the cluster)

Further information about the job submission and an example jobscript can be found on the related page in the wiki: SLURM Job Management (Queueing) System