This web page will no longer be updated. Please use this link for current information.

Usage of the Linux Clusters at DESY Zeuthen

Introduction

There are 4 dedicated parallel clusters (blade centers, Miriquid compute nodes) available for running parallel applications, but you can also run parallel MPI jobs in the SGE farm. The documentation in Batch System Usage applies there.

For discussions and information regarding the usage of the PAX cluster a mailing list has been introduced: <zn-cluster AT desy DOT de>. To get subscribed to that list, send an email to <sympa AT desy DOT de> with the subject: subscribe zn-cluster

Hardware

The batch part consists of four separate partitions that are not interconnected: pax12 (rome) has 16 nodes and HDR Infiniband, pax11 (broadwell) and pax10 (haswell) each consist of 32 compute nodes, connected via a FDR Infiniband network.The legacy system is pax9 (sandybridge), 15 nodes connected by a QDR Infiniband network.

Nodes

All Intel nodes have two CPUs (sockets), while the AMD machines have only one.

Software Environment

The pax machines have a software environment that is slightly different from the normal installation, it includes the OpenHPC software stack and a different version of the module command. To build on any machine in the right environment, run the /project/singularity/images/pax.img image. You can submit your jobs if you run the singularity container on a EL7 WGS like this:

singularity run -B /var/run/munge /project/singularity/images/pax.img

You can also submit your jobs from the machine pax9-00.

Building Applications

Use the 'module' command to first add a compiler implementation and then a version of MPI to your path e.g.:

module add gnu mvapich2

OpenHPC provides the module command from the lmod project. It supports more features then the old environment-modules, including dependent modules, that are shown only after loading the prequisites, e.g. for openmpi you'll have to load the intel module first.

Interactive tests

You can run interactive jobs in Slurm after allocating nodes with salloc, e.g.: salloc -p sandybridge -N 2 -c 2. To get an interactive shell on the allocated machines, use the command srun --pty bash.

OpenMPI

To run an MPI program outside the batch system, you must specify a machinefile listing all the machines and the number of cores your application should run on. A typical machine file looks like this:

pax8a slots=8
pax8b slots=8
pax8c slots=8
pax8d slots=8

The command line would look like this:

/opt/ohpc/pub/mpi/openmpi-gnu/1.10.7/bin/mpirun -np 32 -machinefile ./machinefile  ./program

More information on openmpi is in the openmpi FAQ: http://www.open-mpi.org/faq/

Mvapich2

To use mvapich2, add one of those versions to your path and compile your application with that mpi compiler. Applications built with mvapich2 can use only Infiniband network hardware, so they will work on the pax machines, but not on more than one farm machine or WGS.

The machine file format is different from the one for openmpi, you must list the host name for every core you want to use, e.g. if you want to run four processes, two processes on each of pax89 and pax88:

pax88
pax89
pax88
pax89

The preferred way to run a application with mvapich2 is mpiexec, e.g.:

/opt/ohpc/pub/mpi/mvapich2-intel/2.2/bin/mpiexec -n 4 -machinefile ./machinefile /opt/ohpc/pub/libs/intel/mvapich2/imb/2018.1/bin/IMB-MPI1

Intel MPI

To use Intel MPI, add a compiler module followed by impi. Use the compiler wrappers like 'mpicc' and 'mpif90' for GNU or 'mpiicc' and 'mpiifort' for the Intel compiler. To run the resulting application, set the environment variable like this:

export FI_PROVIDER=verbs

In a Slurm job, please use the prun wrapper to start your application.

Batch System Access

ATTENTION: The PAX is now based on the SLURM scheduling system.

Slurm Commands

The most important commands:

Allocation

Slurm was configured to always schedule complete nodes to each job. The pax machines have hyperthreading enabled, each hardware thread is seen as a CPU core by Slurm, so by default, on a 32 core machine with hyperthreading, 64 MPI processes are assigned. To prevent that, use the option -c 2 for sbatch, salloc or srun.

Parallel Execution

Slurm has integrated execution support for parallel programs, replacing mpirun. To work around slight differences in needed options, use prun instead of srun for starting MPI application. You'll have to load the prun module first.

MPI Support

Before running MPI programs, the LD_LIBRARY_PATH variable must first be set, this is done by loading the right environment module, e.g. module add intel openmpi.

Job scripts

Parameters to slurm can be set on the sbatch command line or starting with a #SBATCH in the script. The most important parameters are:

Be careful with --get-user-env, it will also copy loaded modules to the job.

Time format

The runtime of a job is given as minutes, hours, minutes and seconds (HH:MM:SS) or days and hours (DD-HH). The maximum run time was set to 48 hours.

Examples

An example job script is in slurm-mpi.job

Accounting

The jobs and their resources usage is stored in a database that is used for the fair share part of the scheduler. You can view your account's jobs with the command sacct. With no parameters,only today's jobs are shown, to view all jobs since May 1st, use the command sacct -S 2014-05-01 . To view jobs from other accounts as well, use the --allusers option.

Local Disk Space

Each node has a local directory /scratch with up to 1TB of space. It is cleared automatically at the end of the job.

I/O nodes

Most of the pax10 and pax11 machines have external 1GB/s Ethernet connections to the storage. To allow faster storage access, four machines each in the pax10 and pax11 partitions are equipped with 10GB/s Ethernet instead. To access them, you'll have to request the 10g feature in Slurm:  --constraint=10g*1. That way, the first process, the one executing the job scripts, will run on one of the machines with faster connectivity. All pax12 machines have 10GB/s Ethernet as well.

Partitions and backfilling

The cluster consists of four separate partitions: broadwell (default, alias pax), rome, haswell and sandybridge. Jobs can run on only one type of node. The special partition backfill is used for filling up otherwise empty nodes. Jobs running there are automatically terminated by slurm if another job on the main partition needs the nodes.

SL7 changes

As the versions and paths of the MPI implementations have changed, programs are not compatible between SL6 and SL7. You should rebuild your application on SL7, but you could also try singularity.

The 'module' command was replaced by a different, more powerful implementation called lmod. It doesn't list all available modules, instead it supports dependent modules, e.g. the MPI implementations build with 'gnu7' are shown after module add gnu7.

Running EL6 software using Singularity

It is possible to run software built on EL6 in a Singularity container. This works with mvapich2 binaries by calling singularity in the batch script like this:

mpiexec singularity exec /project/singularity/images/SL6.img yourbinary

However, Mvapich2 2.2 isn't optimized yet for Singularity, so this is slower than running native programs.

For Openmpi, singularity is supported in Openmpi >= 2.1, that's why you'll have to rebuild your program with openmpi3 as installed in the SL6 singularity container:

singularity exec /project/singularity/images/SL6.img /usr/lib64/openmpi-3.0/bin/mpicc yourprog.c -o yourprog.sl6

and in the job script:

module add gnu7 openmpi3 prun
prun singularity exec -B /scratch /project/singularity/images/SL6.img yourprog.sl6

Additional Software

The software installation is based on the OpenHPC project. We provide only a subset of the available software. If you need any of the other available components, send a request to zn-cluster@desy.de

AFS Access

The application binary must be available to all nodes, that's why it should be placed in an AFS or Lustre directory.

Monitoring

Ganglia provides a web monitoring interface. These page is only available from the internal network.

Parallel Batch Machines

Known Issues

1. Openmpi3 has a bug that makes the program hang in certain situations: https://www.mail-archive.com/users@lists.open-mpi.org//msg31839.html Use openmpi instead.

2. openmpi 1.10.x jobs crash on pax10-[28-31]. This is caused by the Mellanox Ethernet cards in these nodes. There are several workarounds:
   1. Use openmpi3 or mvapich2 when using these nodes

   2. Exclude them with the -x pax10-[28-31] option to sbatch.

   3. If you use only the nodes pax10-[28-31] with openmpi, exclude the default Infiniband device: set OMPI_MCA_btl_openib_if_exclude=mlx4_0:1

3. You need to acquire an addressless Kerberos ticket for Slurm to work. This is the default on supported DESY machines. On self-maintained machines like notebooks, simply set noaddresses=true in the file /etc/krb5.conf. To check if your ticket is addressless, call klist -v (Heimdal klist only).

4. The command sbcast cannot be used to copy a file to /scratch, as that is a bind mounted directory. Use /batch/job.${SLURM_JOB_ID}.0/scratch as target.

5. The module command might be unavailable for tcsh login shell users. As workaround, they can run bash -l and use the --get-user-env option in the job.

6. There are some compatibility problems between third-party module files (e.g. Intel 2021) and the module command.

Further documentation

Paralleles Rechnen in Zeuthen - die neuen Cluster , 04/27/10, technical seminar

HPC-Clusters at DESY Zeuthen , 11/22/06, technical seminar