Overview

Access

For all ALICE users

• All ALICE users can access part of the resources on node802 using partition "amd_short".

For MI users

• ALICE users from MI also have access to partition "mem_mi". This partition is only accessible for MI users.

• Jobs submitted to this partition should always get a higher priority than jobs submitted to amd-short. If you notice any issues with this, please contact the ALICE Helpdesk.

Hardware

• The basic hardware configuration of node802 is available here: https://pubappslu.atlassian.net/wiki/spaces/HPCWIKI/pages/37519378/About+ALICE#Hardware-Description

• Hyperthreading is active

• Local node scratch

  • The two 10TB HDDs were combined into a single volume of about 20TB mounted at /scratchdata

  • You can use the local scratch on node802 in the same way that you would use it on other ALICE nodes.

Software

Scientific software stack

• Because this node uses AMD CPUs, there is a new and separate AMD software stack.

  • The main ALICE software stack was compiled with Intel CPUs and should not be used for jobs on node802.

  • New software can be added to the stack by contacting the ALICE Helpdesk.

• It might be necessary to recompile your own software for running on node802.

• This is also the case for conda and Python virtual environment

One option

You can make use of the general scientific software stack which can be accessed by running

Code Block
module load ALICE/default

It is recommend to add this to your batch scripts, too.

• If you want to use software fully optimized for the CPU architecture of the nodes, you have to build the software yourself.

Your own scripts/programmes

• Because this node has a different CPU, it is possible that conda environments or other software that you build on the login nodes are not working if the software is build optimized for CPU architecture.

• In this case, you need to compile such scripts/software as part of a batch or interactive job

• Because a separate software stack is necessary for an AMD machine, you should not compile jobs with GNU-compilers or the likes on the login nodes which are Intel based

• You should always compile such scripts/software as part of your job

• If you do not make any changes you can compile your program on node802

  • One way to do this is to create a short slurm batch job specifically for compiling your software, setting up your conda/Python environments, etc. If you only need to do this once, then there is no need to make this part of your production batch job.

• To get a list of currently available modules in the AMD software stack when you are on a login node, use module load ALICE/AMD and then module avail

  • If you want to go back to the Intel software stack, use module load ALICE/Intel

• For a Slurm batch script that you submit to node802, you can also use module load ALICE/default which will load the correct software stack based on the CPU architecture of the node.

  • You can include module load ALICE/default in all your batch scripts if you like

Your own scripts/programmes

• • Another option is to compile the first time you run

  it

  • your programme as part of a job. In this first job, you copy the compiled program back to your shared storage or home directory. For the next job, you use the already compiled version (see example below).

• You can still use the login nodes for testing/debugging. In this case, you need to compile on the login nodes, run your test and for your job, compile on the compute node again.

Example

Here is an example of how a Slurm batch script could look like for using the node, including a HelloWorld OpenMP program to demonstrate the compiling and use of the local scratch storage.

If you are new to HPC, ALICE or Slurm, have a look at the https://pubappslu.atlassian.net/wiki/spaces/HPCWIKI/pages/5963809 first.

Batch script

#!/bin/bash
#SBATCH --partition=mem_mi
#SBATCH --reservation=mem_mi
#SBATCH --job-name=test_job
#SBATCH --time=0-00:02:00
#SBATCH --output=%x_%j.out
#SBATCH --nodes=1
#SBATCH --ntasks=5
#SBATCH --cpus-per-task=3
#SBATCH --mem=10G
#SBATCH --mail-user="your-email-address"
#SBATCH --mail-type="ALL"

module load ALICE/default
module load OpenMPI/4.0.5-GCC-9.3.0

echo "#### Test started"

# return the name of the node
echo "## Which node is this: $HOSTNAME"

# check the number of cores (ntasks*cpus-per-task)
echo "How many cores do I have access to: ${SLURM_CPUS_ON_NODE}"

# Just to check that the AMD software stack is loaded
echo "Am I loading the from the right module path"
echo ${MODULEPATH%%:*}

# get the current working directory
CWD=$(pwd)

echo "## Where am I: ${CWD}"

# check out the nodes local scratch
echo "## My local scratch space on the node is: ${SCRATCH}"
cd $SCRATCH

echo "## Let us go there: $(pwd)"

# In case the file has already been compiled
# and stored in $CWD, the following six lines
# are not necessary  
echo "## Let us copy the C script to it"
cp $CWD/omp_hello.c $SCRATCH/  
echo "## Is the file there?"
ls -la omp_hello.c
echo "## Now we compile it on the node"
gcc -o omp_hello_amd -fopenmp omp_hello.c

# In case the file is already compiled
# the next four lines would copy it
# and check that it is there:
#echo "## Let us copy the compiled C programme to it"
#cp $CWD/omp_hello_amd $SCRATCH/
#echo "## Is the file there?"
#ls -la omp_hello_amd

echo "## Let us run it"
export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASKS
srun ./omp_hello_amd

# Copy those files back to shared scratch or home
# that should be kept for later.
# Here, it is just the compiled C programme.
# It does not need to be copied back of course
# if it came from shared scratch or home.
echo "## Saving files that should be saved."
cp $SCRATCH/omp_hello_amd $CWD/

echo "## Now that this is done, I want to go home"
cd $CWD
echo "## Good to be back $(pwd)"

echo "#### Test finished"

OpenMP script

Here is the content of the file omp_hello.c from https://computing.llnl.gov/tutorials/openMP/samples/C/omp_hello.c

/******************************************************************************
 * * FILE: omp_hello.c
 * * DESCRIPTION:
 * *   OpenMP Example - Hello World - C/C++ Version
 * *   In this simple example, the master thread forks a parallel region.
 * *   All threads in the team obtain their unique thread number and print it.
 * *   The master thread only prints the total number of threads.  Two OpenMP
 * *   library routines are used to obtain the number of threads and each
 * *   thread's number.
 * * AUTHOR: Blaise Barney  5/99
 * * LAST REVISED: 04/06/05
 * ******************************************************************************/
#include <omp.h>
#include <stdio.h>
#include <stdlib.h>

int main (int argc, char *argv[])
{
int nthreads, tid;

/* Fork a team of threads giving them their own copies of variables */
#pragma omp parallel private(nthreads, tid)
  {

  /* Obtain thread number */
  tid = omp_get_thread_num();
  printf("Hello World from thread = %d\n", tid);

  /* Only master thread does this */
  if (tid == 0)
    {
    nthreads = omp_get_num_threads();
    printf("Number of threads = %d\n", nthreads);
    }

  }  /* All threads join master thread and disband */

}