HPC/Carbon Cluster - Development tools
Compilers and development tools
GNU
Intel
We have supported versions of the following commercial tools, currently installed under /opt/intel
:
- Fortran
- Manual for 10.1
/opt/intel/{fc,fce}/version
- C++
- Manual for 10.1
/opt/intel/{cc,cce}/version
- Intel Math Kernel Library (MKL), includes Cluster Edition
/opt/intel/mkl/version
- VTune analyzer
- Intel Debugger
/opt/intel/{idb,idbe}/version
To access the documentation pages, use firefox on a login node:
/opt/intel/fce/version/doc/Doc_Index.htm
/opt/intel/cce/version/doc/Doc_Index.htm
/opt/intel/mkl/version/doc/Doc_index.htm
Note on binary architectures
CPU architecture labels used by Intel and others can be confusing and have changed over time. Here's a summary of the labels and directory names commonly used:
- 32, ia32, x86
- legacy 32-bit x86 generations up to Pentium 4 and Intel Core
- em64t, Intel-64, x86_64, AMD-64
- 64-bit generations of the Xeon branded CPUs, and the Core 2 generation, mostly compatible with AMD's Opteron
- 64, ia64, Itanium
- Itanium – limited ia32 support, but none for x86_64; not used here.
The cce, fce, and idbe directories above provide the 64-bit em64t versions appropriate for the Xeon processors, with ia32 versions being available under cc, fc, idb to compile and run legacy applications. By default, the 64-bit versions are activated in the user's environment ($PATH
and others). This is done automatically through /etc/profile.d/{iccvars,ifortvars,mklvarsem64t}.{sh,csh}
.
In the future, we will leverage the environment-modules package for version and architecture selection.
Libraries
Intel Math Kernel Library (MKL)
User Notes
- User's Guide – consult Sec. 5 to resolve linker issues, and 32bit vs. 64bit programming
- Reference Manual
- since v10 includes formerly separate Cluster MKL (providing ScaLapack et al.)
- provides BLAS, Lapack, ScaLapack, FFTs (including an FFTW interface)
- Detailed linking instructions
- Link example (from
/opt/cnm/examples
):
# Sequential: -lmkl_intel_lp64 -lmkl_sequential -lmkl_lapack -lmkl_em64t -lmkl_core -lpthread
# alternative: -Wl,--start-group -lmkl_intel_lp64 -lmkl_sequential -lmkl_lapack -lmkl_em64t -lmkl_core -Wl,--end-group \ -lpthread
# OpenMP multi-threaded: -Wl,--start-group -lmkl_intel_lp64 -lmkl_intel_thread -lmkl_lapack -lmkl_em64t -lmkl_core -Wl,--end-group \ -lguide -lpthread
- The FFTW interface of the MKL can be linked by prefixing the above options with one of:
-lfftw2xc_intel -lfftw2xf_intel -lfftw3xc_intel -lfftw3xf_intel
Choose the appropriate version between bindings for FFTW v2 vs. v3 and C vs. Fortran.
- Avoid specifying absolute library paths in your Makefile unless you wish to select a specific version. The user environment is normally correctly set up. To force the library version current at the time of writing, add as first flag to the linker options:
-L$MKLROOT/lib/em64t
(automatically set in $LD_LIBRARY_PATH
).
- see HPC/Carbon_Cluster_-_Getting_started#Example_job_file for controlling OpenMP execution.
VTK
PETSc
- TODO: Portable, Extensible Toolkit for Scientific Computation
MPI
OpenMPI
This is the primary and recommended MPI version. It supports Ethernet and Infiniband interconnect.
- See also: FAQ
MPICH/MPIC2
TBD
Intel MPI
Resource Management
- MOAB cluster suite
- Documentation overview
- Torque
- Gold allocation manager
(description of overlap/differences)