-
Thomas Jahns authoredThomas Jahns authored
README.build 8.85 KiB
How to build yaxt
This README is divided into the following topics of increasing specificity.
1. After checkout
2. General advice for running configure
3. Requirements
4. Specific machine setups
4.1 Debian GNU/Linux Squeeze 6 x86_64 with OpenMPI 1.4.3
4.2 IBM AIX 6.1, POWER archictecture
4.3 Redhat RHEL6 x86_64 with Intel MPI 4.1.3.049 and compilers 14.0.3
4.4 Cray CCE
4.5 IBM BlueGene/Q with xlc/xlf
1. After checkout
To initialize the autotools, execute
$ scripts/reconfigure
or try
$ libtoolize
$ autoreconf -i
(but on macOS glibtoolize might be needed instead of libtoolize).
This is not required if you have obtained a regular, packaged source
distribution of yaxt, but only in case you have directly accessed the
repository.
2. General advice for running configure
To adapt yaxt to your MPI system, yaxt needs to be given two to four pieces of
information
a) how to invoke the C compiler and compile and link C programs
b) how to invoke the Fortran compiler and compile and link Fortran programs
c) how to run compiled MPI programs (mpirun and mpiexec will be
tried automatically)
d) how to produce C code for C<->Fortran interoperability, see cfortran.doc
for a number of compilers this will be found automatically
For a) and b) respectively you can either
- set CC and FC to the corresponding MPI compiler wrapper, or
- set CC and FC to the compiler driver (e.g. CC=gcc FC=gfortran) and use
some of the arguments MPIROOT MPI_C_INCLUDE MPI_FC_MOD,
MPI_C_LIB and MPI_FC_LIB to set the necessary compiler flags to build
MPI programs.
For c) the configure-argument MPI_LAUNCH needs to be set according to
your MPI installation. For mpich/openmpi style environments, mpirun is
fine and usually detected automatically. For IBM PE an adapter that's
provided as util/mpi_launch_poe in the yaxt source must be used.
For d) the following is used as an option in CPPFLAGS for different
compilers automatically:
* GCC gfortran, PGI pgfortran, Cray and Intel compilers: -DgFortran
* NAG Fortran compiler: -DNAGf90Fortran
* IBM xlc needs -Dappendus if -qextname is used for the Fortran part.
Additionally the following considerations can be useful later on:
* It is possible to build yaxt out-of-source, i.e. to call configure
from a directory different from the place the source was put in. If
one has e.g. unpacked yaxt in $HOME/src/yaxt and wants to do the build in $HOME/build/yaxt, one would run the following commands:
$ cd $HOME/build/yaxt
$ ../../src/yaxt/configure [OPTIONS-NEEDED-FOR-PLATFORM]
See descriptions under 4. for the options most appropriate for the
system in question.
*
3. Requirements
Yaxt requires the following software environment to function
- a C compiler with C99 language level support
- a Fortran compiler with F2003 C interoperability and
set (by flags, wrappers or whatever) to preprocess .f90 files.
- an MPI library working with both of the above compilers
4. Specific machine setups
The following system setups are regularly being tested:
4.1 Debian GNU/Linux Squeeze 6 x86_64 with OpenMPI 1.4.3
The following packages need to be installed:
libopenmpi-dev
The following commands should produce a working install:
$ ./configure CC=mpicc FC=mpif90 FCFLAGS=-cpp
$ make
$ make check
$ make install
equivalently one can use:
$ ./configure CC=gcc FC=gfortran MPIROOT=/usr/lib/openmpi \
MPI_FC_LIB='-L/usr/lib/openmpi -lmpi_f90 -lmpi_f77' \
FCFLAGS=-cpp
for the configure run.
4.2 IBM AIX 6.1, POWER archictecture
The following is used to build the library and test programs for blizzard,
a POWER6 system running with AIX and Parallel Environment (PE).
* Shared library compiler wrappers (smartmpcc_r and smartmpxlf_r) are used
(instead of mpcc_r and mpxlf_r), this requires users to
extend the PATH variable by /sw/aix61/smartmpxlf/bin
otherwise one has to use the --disable-shared option.
IBM PE compiler frontends would otherwise insert start-up code into the
shared libraries that make a program using the libraries fail.
* The options --host=powerpc-ibm-aix6.1.0.0 and
host_alias=powerpc-ibm-aix6.1.0.0 put configure in cross-compilation
mode since PE executables cannot easily be directly executed.
* The common optimization options (-O3 -q64 -qarch=pwr6 -qtune=pwr6
-qipa -qinline -qhot -qxflag=nvectver -qxflag=nsmine -qstrict) are typically
used to achieve a good optimization level but not strictly required and
should be changed to disable optimization when debugging
(Recommendation: -O0 -g -qfullpath -q64 -qarch=pwr6 -qtune=pwr6).
* The archive must be given the -X32_64 option such that static library
archives can be created regardless of the setting of the OBJECT_MODE
environment variable.
* TLS storage should be enabled for the C compiler via the -qtls=initial-exec option (or -qtls=global-dynamic if plugins with TLS data
are needed).
* A wrapper to make mpirun-style arguments compatible with the PE program
starter (poe) must be used to enable 'make check'.
* It's recommended to use /bin/bash instead of /bin/sh for running configure
because of speed reasons.
$ /bin/bash ./configure CC=smartmpcc_r \
CFLAGS='-O3 -q64 -qarch=pwr6 -qtune=pwr6 -qipa -qinline -qhot -qxflag=nvectver -qxflag=nsmine -qstrict -qlanglvl=extc99 -qtls=initial-exec' \
FC=smartmpxlf_r \
FCFLAGS='-O3 -q64 -qarch=pwr6 -qtune=pwr6 -qipa -qinline -qhot -qxflag=nvecvter -qxflag=nsmine -qstrict -qessl -qextname -qsuffix=cpp=f90 -qzerosize -qfloat=fltint -qhalt=w -Q' \
CPPFLAGS=-Dextname \
LDFLAGS="-O3 -q64 -qarch=pwr6 -qtune=pwr6 -qipa -qinline -qhot -qxflag=nvecvter -qxflag=nsmine -qstrict -Wl,-brtl" \
--host=powerpc-ibm-aix6.1.0.0 host_alias=powerpc-ibm-aix6.1.0.0 \
AR='ar -X32_64' \
MPI_LAUNCH="$(pwd)/util/mpi_launch_poe" \
CONFIG_SHELL=/bin/bash
4.3 Redhat RHEL6 x86_64 with Intel MPI 4.1.3.049 and compilers 14.0.3
To use the Intel compilers with Intel MPI, slightly differently named
wrappers for the C and Fortran compilers can be used, namely mpiicc
and mpiifort:
$ ./configure CC=mpiicc \
CFLAGS='-O3 -fno-math-errno -march=native -mtune=native' \
FC=mpiifort \
FCFLAGS='-O3 -heap-arrays -fpp -march=native -mtune=native'
4.4 Cray CCE
The Cray compiler is a cross-compiler because it builds binaries that
cannot (fully functionally) run on the interactively usable frontend
machines. The aprun program starter is functionally similar to mpirun
on more conventional platforms but requires that the requested
resources are available to the user. For similar reasons the Cray
compiler builds static executables/libraries only.
Assuming the Cray system uses PBS, its MPI launcher aprun cannot be
executed directly but must be started from a PBS job. We provide a
wrapper for this purpose in util/pbs_aprun that acts like mpirun but
creates a job in the background and forwards its output and error code
to the frontend machine. On systems with a different batch scheduler,
a different/adapted wrapper might be necessary. We would be interested
to distribute such wrappers with yaxt.
To build yaxt on a Cray system with PBS Pro the following configure
was used successfully (note the --host/--build options to switch to
cross compilation mode):
$ ./configure CC=cc FC=ftn FCFLAGS= MPI_LAUNCH=$(pwd)/util/pbs_aprun \
--host=x86_64-cray-linux-gnu --build=x86_64-unknown-linux-gnu \
--disable-shared \
build_alias=x86_64-unknown-linux-gnu host_alias=x86_64-cray-linux-gnu
4.5 IBM BlueGene/Q with xlc/xlf
On BlueGene similar considerations as for aprun on Cray machines are
true for the runjob starter program: one must have reserved
corresponding node resources before starting the job. Also command
syntax is slightly different from mpirun so that a wrapper must be
used (to be found in the util directory of the yaxt source). On
systems that use IBM Tivoli Load Leveler for this purpose, one can run
the configure and make check in jobs. The following job files where
used on one system for this purpose:
For configure:
#! /bin/bash# @ job_name = yaxt-bgq-configure
# @ comment = "YAXT requires having a working MPI-starter for configure and tests"
# @ error = LOG.$(job_name).$(jobid).out
# @ output = LOG.$(job_name).$(jobid).out
# @ environment = COPY_ALL
# @ wall_clock_limit = 00:15:00
# @ notification = error
# @ notify_user = some@valid.address.org
# @ job_type = bluegene
# @ bg_size = 32
# @ queue
exec ./configure \
--disable-shared \
FC=mpixlf2003_r \
FCFLAGS='-O2 -g -qarch=qp' \
CC="mpixlc_r -qlanglvl=extc99" \
CFLAGS='-O2 -g -qarch=qp' \
MPI_LAUNCH="$(readlink -f ./util/mpi_launch_bgrunjob)" \
LDFLAGS='-g'
For make check:
#! /bin/bash
# @ job_name = yaxt-bgq-make-check
# @ comment = "YAXT requires having a working MPI-starter for configure and tests"
# @ error = LOG.$(job_name).$(jobid).out
# @ output = LOG.$(job_name).$(jobid).out
# @ environment = COPY_ALL
# @ wall_clock_limit = 00:15:00
# @ notification = error
# @ notify_user = some@valid.address.org
# @ job_type = bluegene
# @ bg_size = 32
# @ queue
make check