Application development environment
ARCHER2 runs on the Cray Linux Environment (a version of SUSE Linux), and provides a development environment which includes:
- Software modules via a standard module framework
- Three different compiler environments (AMD, Cray, and GNU)
- MPI, OpenMP, and SHMEM
- Scientific and numerical libraries
- Parallel Python and R
- Parallel debugging and profiling
- Singularity containers
Access to particular software, and particular versions, is managed by a standard TCL module framework. Most software is available via standard software modules and the different programming environments are available via module collections.
You can see what programming environments are available with:
auser@uan01:~> module savelist Named collection list: 1) PrgEnv-aocc 2) PrgEnv-cray 3) PrgEnv-gnu
Other software modules can be listed with
auser@uan01:~> module avail ------------------------------- /opt/cray/pe/perftools/20.09.0/modulefiles -------------------------------- perftools perftools-lite-events perftools-lite-hbm perftools-nwpc perftools-lite perftools-lite-gpu perftools-lite-loops perftools-preload ---------------------------------- /opt/cray/pe/craype/2.7.0/modulefiles ---------------------------------- craype-hugepages1G craype-hugepages8M craype-hugepages128M craype-network-ofi craype-hugepages2G craype-hugepages16M craype-hugepages256M craype-network-slingshot10 craype-hugepages2M craype-hugepages32M craype-hugepages512M craype-x86-rome craype-hugepages4M craype-hugepages64M craype-network-none ------------------------------------- /usr/local/Modules/modulefiles -------------------------------------- dot module-git module-info modules null use.own -------------------------------------- /opt/cray/pe/cpe-prgenv/7.0.0 -------------------------------------- cpe-aocc cpe-cray cpe-gnu -------------------------------------------- /opt/modulefiles --------------------------------------------- aocc/220.127.116.11(default) cray-R/18.104.22.168(default) gcc/8.1.0 gcc/9.3.0 gcc/10.1.0(default) ---------------------------------------- /opt/cray/pe/modulefiles ----------------------------------------- atp/3.7.4(default) cray-mpich-abi/8.0.15 craype-dl-plugin-py3/20.06.1(default) cce/10.0.3(default) cray-mpich-ucx/8.0.15 craype/2.7.0(default) cray-ccdb/4.7.1(default) cray-mpich/8.0.15(default) craypkg-gen/1.3.10(default) cray-cti/2.7.3(default) cray-netcdf-hdf5parallel/22.214.171.124 gdb4hpc/4.7.3(default) cray-dsmml/0.1.2(default) cray-netcdf/126.96.36.199 iobuf/2.0.10(default) cray-fftw/188.8.131.52(default) cray-openshmemx/11.1.1(default) papi/184.108.40.206(default) cray-ga/220.127.116.11 cray-parallel-netcdf/18.104.22.168 perftools-base/20.09.0(default) cray-hdf5-parallel/22.214.171.124 cray-pmi-lib/6.0.6(default) valgrind4hpc/2.7.2(default) cray-hdf5/126.96.36.199 cray-pmi/6.0.6(default) cray-libsci/20.08.1.2(default) cray-python/188.8.131.52(default)
A full discussion of the module system is available in the Software environment section.
A consistent set of modules is loaded on login to the machine (currently
PrgEnv-cray, see below). Developing applications then means selecting
and loading the appropriate set of modules before starting work.
This section is aimed at code developers and will concentrate on the compilation environment and building libraries and executables, and specifically parallel executables. Other topics such as Python and Containers are covered in more detail in separate sections of the documentation.
ARCHER2 supports common revision control software such as
Standard GNU autoconf tools are available, along with
make (which is
GNU Make). Versions of
cmake are available.
Some of these tools are part of the system software, and
typically reside in
/usr/bin, while others are provided as part of the
module system. Some tools may be available in different versions via
/usr/bin and via the module system.
There are three different compiler environments available on ARCHER2: AMD (AOCC), Cray (CCE), and GNU (GCC). The current compiler suite is selected via the programming environment, while the specific compiler versions are determined by the relevant compiler module. A summary is:
|Suite name||Module||Programming environment collection|
For example, at login, the default set of modules are:
Currently Loaded Modulefiles: 1) cpe-cray 7) cray-dsmml/0.1.2(default) 2) cce/10.0.3(default) 8) perftools-base/20.09.0(default) 3) craype/2.7.0(default) 9) xpmem/2.2.35-184.108.40.206_1.3__gd50fabf.shasta(default) 4) craype-x86-rome 10) cray-mpich/8.0.15(default) 5) libfabric/220.127.116.11.233(default) 11) cray-libsci/20.08.1.2(default) 6) craype-network-ofi
from which we see the default programming environment is Cray (indicated
cpe-cray (at 1 in the list above) and the default compiler module
cce/10.0.3 (at 2 in the list above). The programming environment
will give access to a consistent set of compiler, MPI library via
cray-mpich (at 10), and other libraries e.g.,
cray-libsci (at 11 in
the list above) infrastructure.
Within a given programming environment, it is possible to swap to a different compiler version by swapping the relevant compiler module.
To ensure consistent behaviour, compilation of C, C++, and Fortran
source code should then take place using the appropriate compiler
ftn, respectively. The wrapper will
automatically call the relevant underlying compiler and add the
appropriate include directories and library locations to the invocation.
This typically eliminates the need to specify this additional
information explicitly in the configuration stage. To see the details of
the exact compiler invocation use the
-craype-verbose flag to the
The default link time behaviour is also related to the current programming environment. See the section below on Linking and libraries.
Users should not, in general, invoke specific compilers at compile/link
stages. In particular,
gcc, which may default to
should not be used. The compiler wrappers
be used via the appropriate module. Other common MPI compiler wrappers
mpicc should also be replaced by the relevant wrapper
mpicc etc are not available).
Always use the compiler wrappers
ftn and not a
specific compiler invocation. This will ensure consistent compile/link
Compiler man pages and help
Further information on both the compiler wrappers, and the individual
compilers themselves are available via the command line, and via
man pages. The
man page for the compiler wrappers is common
to all programming environments, while the
man page for individual
compilers depends on the currently loaded programming environment. The
following table summarises options for obtaining information on the
compiler and compile options:
You can also pass the
--help option to any of the compilers or
wrappers to get a summary of how to use them. The Cray Fortran
ftn --craype-help to access the help options.
There are no
man pages for the AOCC compilers at the moment.
Cray C/C++ is based on Clang and therefore
supports similar options to clang/gcc (
man clang is in fact equivalent
clang --help will produce a full summary
of options with Cray-specific options marked "Cray". The
page concentrates on these Cray extensions to the
clang front end and
does not provide an exhaustive description of all
Cray Fortran is not based on Flang and so takes different options
Executables on ARCHER2 link dynamically, and the Cray Programming Environment does not currently support static linking. This is in contrast to ARCHER where the default was to build statically.
If you attempt to link statically, you will see errors similar to:
/usr/bin/ld: cannot find -lpmi /usr/bin/ld: cannot find -lpmi2 collect2: error: ld returned 1 exit status
The compiler wrapper scripts on ARCHER link runtime libraries in using
runpath by default. This means that the paths to the runtime
libraries are encoded into the executable so you do not need to load the
compiler environment in your job submission scripts.
Which compiler environment?
If you are unsure which compiler you should choose, we suggest the
starting point should be the GNU compiler collection (GCC,
PrgEnv-gnu); this is perhaps the most commonly used by code
developers, particularly in the open source software domain. A portable,
standard-conforming code should (in principle) compile in any of the
three programming environments.
For users requiring specific compiler features, such as co-array Fortran, the recommended starting point would be Cray. The following sections provide further details of the different programming environments.
Intel compilers are not available on ARCHER2.
AMD Optimizing C/C++ Compiler (AOCC)
The AMD Optimizing C/++ Compiler (AOCC) is a clang-based optimising compiler. AOCC (despite its name) includes a flang-based Fortran compiler.
Switch the the AOCC programming environment via
$ module restore PrgEnv-aocc
Further details on AOCC will appear here as they become available.
AOCC reference material
- AMD website
Cray compiler environment (CCE)
The Cray compiler environment (CCE) is the default compiler at the point of login. CCE supports C/C++ (along with unified parallel C UPC), and Fortran (including co-array Fortran). Support for OpenMP parallelism is available for both C/C++ and Fortran (currently OpenMP 4.5, with a number of exceptions).
The Cray C/C++ compiler is based on a clang front end, and so compiler
options are similar to those for gcc/clang. However, the Fortran
compiler remains based around Cray-specific options. Be sure to separate
C/C++ compiler options and Fortran compiler options (typically
FFLAGS) if compiling mixed C/Fortran applications.
Switch the the Cray programming environment via
$ module restore PrgEnv-cray
Useful CCE C/C++ options
When using the compiler wrappers
CC, some of the following
options may be
Language, warning, Debugging options:
||Optimisation levels: -O0, -O1, -O2, -O3, -Ofast|
||Floating point maths optimisations levels 0-4 |
||Link time optimisation|
||Compile OpenMP (default is off)|
||Display verbose output from compiler stages|
c11plus GNU extensions (likewise
c++14plus GNU extensions). See https://gcc.gnu.org/onlinedocs/gcc-4.8.2/gcc/C-Extensions.html
-ffp=3is implied by
Useful CCE Fortran options
Language, Warning, Debugging options:
||Message level (default
||Optimisation levels: -O0 to -O3 (default -O2)|
||Floating point maths optimisations levels 0-3|
||Compile OpenMP (default is
||Display verbose output from compiler stages|
GNU compiler collection (GCC)
The commonly used open source GNU compiler collection is available and provides C/C++ and Fortran compilers.
The GNU compiler collection is loaded by switching to the GNU programming environment:
$ module restore PrgEnv-gnu
gcc/8.1.0 module is available on ARCHER2 but cannot be used as the
supporting scientific and system libraries are not available. You should
not use this version of GCC.
If you want to use GCC version 10 or greater to compile Fortran code,
with the old MPI interfaces (i.e.
use mpi or
INCLUDE 'mpif.h') you
must add the
-fallow-argument-mismatch option (or equivalent) when compiling
otherwise you will see compile errors associated with MPI functions.
The reason for this is that past versions of
gfortran have allowed
mismatched arguments to external procedures (e.g., where an explicit
interface is not available). This is often the case for MPI routines
using the old MPI interfaces where arrays of different types are passed
to, for example,
MPI_Send(). This will now generate an error as not
standard conforming. The
-fallow-argument-mismatch option is used
to reduce the error to a warning. The same effect may be achieved via
If you use the Fortran 2008 MPI interface (i.e.
use mpi_f08) then you
should not need to add this option.
Fortran language MPI bindings are described in more detail at in the MPI Standard documentation.
Useful Gnu Fortran options
||Default is gnu|
||Allow mismatched procedure arguments. This argument is required for compiling MPI Fortran code with GCC version 10 or greater if you are using the older MPI interfaces (see warning above)|
||Use runtime checking of array indices|
||Compile OpenMP (default is no OpenMP)|
||Display verbose output from compiler stages|
-std may be one of
f2018. The default option
-std=gnu is the latest Fortran standard
plus gnu extensions.
Past versions of
gfortran have allowed mismatched arguments to
external procedures (e.g., where an explicit interface is not
available). This is often the case for MPI routines where arrays of
different types are passed to
MPI_Send() and so on. This will now
generate an error as not standard conforming. Use
-fallow-argument-mismatch to reduce the error to a warning. The same
effect may be achieved via
Message passing interface (MPI)
HPE Cray MPICH
HPE Cray provide, as standard, an MPICH implementation of the message passing interface which is specifically optimised for the ARCHER2 network. The current implementation supports MPI standard version 3.1.
The HPE Cray MPICH implementation is linked into software by default when
compiling using the standard wrapper scripts:
MPI reference material
MPI standard documents: https://www.mpi-forum.org/docs/
Linking and libraries
Linking to libraries is performed dynamically on ARCHER2. One can use
-craype-verbose flag to the compiler wrapper to check exactly what
linker arguments are invoked. The compiler wrapper scripts encode the
paths to the programming environment system libraries using RUNPATH.
This ensures that the executable can find the correct runtime
libraries without the matching software modules loaded.
The library RUNPATH associated with an executable can be inspected via, e.g.,
$ readelf -d ./a.out
a.out for the name of the executable you are querying).
Commonly used libraries
Modules with names prefixed by
cray- are provided by HPE Cray, and are
supported to be consistent with any of the programming environments and
associated compilers. These modules should be the first choice for
access to software libraries if available.
More information on the different software libraries on ARCHER2 can be found in the Software libraries section of the user guide.
Build instructions for software on ARCHER2
The ARCHER2 CSE team at EPCC and other contributors provide build configurations ando instructions for a range of research software, software libraries and tools on a variety of HPC systems (including ARCHER2) in a public Github repository. See:
The repository always welcomes contributions from the ARCHER2 user community.
Support for building software on ARCHER2
If you run into issues building software on ARCHER2 or the software you require is not available then please contact the ARCHER2 Service Desk with any questions you have.