Skip to content

Frequently asked questions - Software

I cannot find my software — what are my options?

Work through the following steps before submitting a support request.

1. Search the full module tree

LUNARC uses a hierarchical module system where not all software is visible by default. Use module spider to search the entire tree:

module spider <packagename>

The search is case-insensitive. If the package is installed, module spider will find it and show you which prerequisite modules to load first. See Using installed software for details.

2. Consider self-service alternatives

If the software is not installed as a module, you may be able to set it up yourself without waiting for a support request:

Conda / pip (Python packages) Install packages into a personal conda environment or with pip --user. This works well for Python libraries and tools. See the Python guide for how to create and use conda environments on COSMOS.

Containers (Apptainer) If the software has complex dependencies or requires a different OS environment, run it in a container. Many research tools have pre-built containers on Docker Hub or Singularity Hub that can be pulled and run directly on COSMOS. See the containers guide for details.

3. Request a module installation

If the software is not available as a module and the self-service options above are not suitable, submit a support request via SUPR. Include the following to help the team process your request efficiently:

  • Software name and version — the exact version you need
  • Where to get it — a link to the project website, source repository, or documentation
  • Why it is needed — a brief description of your use case
  • Dependencies — any specific libraries, compilers, or toolchains it requires, if known
  • Urgency — if you have a deadline, mention it

LUNARC installs software using EasyBuild. If an EasyBuild recipe already exists for your software, installation is usually straightforward. For software without an existing recipe, the process takes longer.

Loading a module unloads prerequisites from a previously loaded module

Software modules can have conflicting prerequisites. When you load a second module whose prerequisites clash with the first, SLURM will purge the conflicting ones — potentially breaking the earlier module.

Options to resolve this:

  • Check whether different versions of the conflicting modules share common prerequisites. We often provide the same software built against different toolchains for this reason.
  • Use a separate terminal for each module. This works when the modules produce or consume files rather than interacting directly. The LUNARC HPC Desktop makes it easy to run multiple terminals in one session.
  • If neither option works, contact the LUNARC helpdesk to discuss your situation.

My application fails because it can't find libgfortran.so.3

error while loading shared libraries: libgfortran.so.3: cannot
open shared object file: No such file or directory

This means your software was compiled with a deprecated version of GCC. Recompile it on COSMOS using a newer GCC version — preferably one provided as a module.

I need LAPACK and BLAS — they used to be in /usr/lib64

The libraries in /usr/lib64 are not highly optimised. LUNARC provides:

  • OpenBLAS — included in the foss toolchain
  • MKL — included in the intel, iomkl, and gimkl toolchains

Both deliver excellent performance. If you use OpenBLAS in an MPI job, see below for thread control.

LUNARC's MKL modules set the MKLROOT environment variable automatically. For the correct linker flags, use the Intel MKL Link Line Advisor.

My MPI code using OpenBLAS runs slowly / does not perform as expected

OpenBLAS in the foss toolchains is built with thread support and will spawn one thread per allocated core by default. For pure MPI jobs this is usually undesirable. Set OMP_NUM_THREADS to control it:

export OMP_NUM_THREADS=1

See also: compiling MPI code with a toolchain.

I want to run VASP on COSMOS but it doesn't work

VASP requires a valid license registered with the VASP developers in Vienna. Being named in a SUPR VASP group does not grant access.

To request access:

  1. Register your license with the VASP developers
  2. Contact the LUNARC helpdesk with:
    • Your LUNARC user ID
    • Name and department of the license holder
    • License number
    • Your name as it appears on the license
    • Your email address as it appears on the license
    • Whether you need VASP 5 or VASP 6

Note: verification with the VASP developers typically takes a few days but can take longer. VASP executables are built with OpenMPI and must be launched with mpirun. See the MPI example scripts for reference.

I need access to Amber

Amber 22 is licensed software. Review the license on the Amber website, paying particular attention to the Assignment Restrictions.

Submit a support request confirming that you accept the license terms. Your account will then be enabled for the Amber module.

What profiling tools are available on COSMOS?

Several profiling and performance analysis tools are installed as modules. Use module spider <name> to find available versions and prerequisites.

Tool Module Best for
gprof Part of GCC — no module needed Basic CPU profiling of serial and threaded code
Valgrind Valgrind Memory error detection, cache profiling, call graphs
gperftools gperftools Low-overhead CPU and heap profiling (Google Performance Tools)
Score-P Score-P MPI and OpenMP performance analysis; produces traces for Scalasca and Vampir

gprof (GCC)

Compile with -pg to enable profiling instrumentation, run your program normally, then analyse the gmon.out file it produces:

gcc -O2 -pg -o myprogram myprogram.c
./myprogram
gprof myprogram gmon.out > profile.txt

Valgrind

Valgrind runs your program in an instrumented environment. The default tool (memcheck) detects memory errors; callgrind profiles CPU usage:

module load Valgrind
valgrind ./myprogram                        # memory checking
valgrind --tool=callgrind ./myprogram       # CPU profiling

Score-P (MPI / OpenMP)

Score-P instruments MPI and OpenMP programs to produce performance traces. Prefix your compile commands with scorep:

module load Score-P
scorep mpicc -O2 -o myprogram myprogram.c

Run the instrumented binary via your normal job script. Score-P writes trace data to a scorep-* directory for analysis with tools such as Scalasca or Vampir.

My MPI application doesn't launch

COSMOS supports two MPI libraries with different launchers:

Library Launcher
OpenMPI (foss toolchain) mpirun
Intel MPI (intel toolchain) srun

Using the wrong launcher for the toolchain your application was built with is the most common cause of this issue. See the toolchain guide for details.

Author: Joachim Hein (LUNARC)

Last Updated: 2024-11-14