INSTALL_parallel

	    Installation instructions for Parallel HDF5
	    -------------------------------------------

0. Use Build Scripts
--------------------
The HDF Group is accumulating build scripts to handle building parallel HDF5
on various platforms (Cray, IBM, SGI, etc...). These scripts are being 
maintained and updated continuously for current and future systems. The reader
is strongly encouraged to consult the repository at,

https://github.com/HDFGroup/build_hdf5

for building parallel HDF5 on these system. All contributions, additions
and fixes to the repository are welcomed and encouraged.


1. Overview
-----------
This file contains instructions for the installation of parallel HDF5 (PHDF5).
It is assumed that you are familiar with the general installation steps as
described in the INSTALL file.  Get familiar with that file before trying
the parallel HDF5 installation.

The remaining of this section explains the requirements to run PHDF5.
Section 2 shows quick instructions for some well know systems.  Section 3
explains the details of the installation steps. Section 4 shows some details
of running the parallel test suites.


1.1. Requirements
-----------------
PHDF5 requires an MPI compiler with MPI-IO support and a POSIX compliant
(Ref. 1) parallel file system.  If you don't know yet, you should first consult
with your system support staff of information how to compile an MPI program,
how to run an MPI application, and how to access the parallel file system.
There are sample MPI-IO C and Fortran programs in the appendix section of
"Sample programs".  You can use them to run simple tests of your MPI compilers
and the parallel file system.


1.2. Further Help
-----------------

For help with installing, questions can be posted to the HDF Forum or sent to the HDF Helpdesk:

   HDF Forum:     https://forum.hdfgroup.org/
   HDF Helpdesk:  https://help.hdfgroup.org/

In your mail, please include the output of "uname -a". If you have run the
"configure" command, attach the output of the command and the content of
the file "config.log".


2. Quick Instruction for known systems
--------------------------------------
The following shows particular steps to run the parallel HDF5 configure for
a few machines we've tested. If your particular platform is not shown or
somehow the steps do not work for yours, please go to the next section for
more detailed explanations.


2.1. Know parallel compilers
----------------------------
HDF5 knows several parallel compilers: mpicc, hcc, mpcc, mpcc_r.  To build
parallel HDF5 with one of the above, just set CC as it and configure.

    $ CC=/usr/local/mpi/bin/mpicc ./configure --enable-parallel --prefix=<install-directory>
    $ make		# build the library
    $ make check	# verify the correctness
			# Read the Details section about parallel tests.
    $ make install


2.2. Linux 2.4 and greater
--------------------------
Be sure that your installation of MPICH was configured with the following
configuration command-line option:

    -cflags="-D_LARGEFILE_SOURCE -D_LARGEFILE64_SOURCE -D_FILE_OFFSET_BITS=64"

This allows for >2GB sized files on Linux systems and is only available with
Linux kernels 2.4 and greater.


2.3. Hopper (Cray XE6) (for v1.8 and later)
-------------------------

The following steps are for building HDF5 for the Hopper compute
nodes.  They would probably work for other Cray systems but have
not been verified.

Obtain the HDF5 source code:
    https://support.hdfgroup.org/downloads/index.html

The entire build process should be done on a MOM node in an interactive allocation and on a file system accessible by all compute nodes.
Request an interactive allocation with qsub:
qsub -I -q debug -l mppwidth=8

- create a build directory build-hdf5:
  mkdir build-hdf5; cd build-hdf5/

- configure HDF5:
  RUNSERIAL="aprun -q -n 1" RUNPARALLEL="aprun -q -n 6" FC=ftn CC=cc /path/to/source/configure --enable-fortran --enable-parallel --disable-shared

  RUNSERIAL and RUNPARALLEL tell the library how it should launch programs that are part of the build procedure.

- Compile HDF5: 
  gmake

- Check HDF5
  gmake check

- Install HDF5
  gmake install

The build will be in build-hdf5/hdf5/ (or whatever you specify in --prefix).
To compile other HDF5 applications use the wrappers created by the build (build-hdf5/hdf5/bin/h5pcc or h5fc)


3. Detail explanation
---------------------

3.1. Installation steps (Uni/Multiple processes modes)
-----------------------
During the step of configure, you must be running in the uni-process mode.
If multiple processes are doing the configure simultaneously, they will
incur errors.

In the build step (make), it depends on your make command whether it can
run correctly in multiple processes mode.  If you are not sure, you should
try running it in uni-process mode.

In the test step (make check), if your system can control number of processes
running in the MPI application, you can just use "make check".  But if your
system (e.g., IBM SP) has a fixed number of processes for each batch run,
you need to do the serial tests by "make check-s", requesting 1 process and
then do the parallel tests by "make check-p", requesting n processes.

Lastly, "make install" should be run in the uni-process mode.


3.2. Configure details
----------------------
The HDF5 library can be configured to use MPI and MPI-IO for parallelism on
a distributed multi-processor system. The easiest way to do this is to have
a properly installed parallel compiler (e.g., MPICH's mpicc or IBM's mpcc_r)
and supply the compiler name as the value of the CC environment variable.
For examples,

    $ CC=mpcc_r ./configure --enable-parallel
    $ CC=/usr/local/mpi/bin/mpicc ./configure --enable-parallel

If a parallel library is being built then configure attempts to determine how
to run a parallel application on one processor and on many processors. If the
compiler is `mpicc' and the user hasn't specified values for RUNSERIAL and
RUNPARALLEL then configure chooses `mpiexec' from the same directory as `mpicc':

    RUNSERIAL:    mpiexec -n 1
    RUNPARALLEL:  mpiexec -n $${NPROCS:=6}

The `$${NPROCS:=6}' will be substituted with the value of the NPROCS
environment variable at the time `make check' is run (or the value 6).

Note that some MPI implementations (e.g. OpenMPI 4.0) disallow oversubscribing
nodes by default so you'll have to either set NPROCS equal to the number of
processors available (or fewer) or redefine RUNPARALLEL with appropriate
flag(s) (--oversubscribe in OpenMPI).

4. Parallel test suite
----------------------
The testpar/ directory contains tests for Parallel HDF5 and MPI-IO.  Here are
some notes about some of the tests.

The t_mpi tests the basic functionalities of some MPI-IO features used by
Parallel HDF5.  It usually exits with non-zero code if a required MPI-IO
feature does not succeed as expected.  One exception is the testing of
accessing files larger than 2GB.  If the underlying filesystem or if the
MPI-IO library fails to handle file sizes larger than 2GB, the test will
print informational messages stating the failure but will not exit with
non-zero code.  Failure to support file size greater than 2GB is not a fatal
error for HDF5 because HDF5 can use other file-drivers such as families of
files to bypass the file size limit.

The t_cache does many small sized I/O requests and may not run well in a
slow file system such as NFS disk. If it takes a long time to run it, try
set the environment variable $HDF5_PARAPREFIX to a file system more suitable
for MPI-IO requests before running t_cache.

By default, the parallel tests use the current directory as the test directory.
This can be changed by the environment variable $HDF5_PARAPREFIX.  For example,
if the tests should use directory /PFS/user/me, do
    HDF5_PARAPREFIX=/PFS/user/me
    export HDF5_PARAPREFIX
    make check

(In some batch job system, you many need to hardset HDF5_PARAPREFIX in the
shell initial files like .profile, .cshrc, etc.)


Reference
---------
1. POSIX Compliant. A good explanation is by Donald Lewin,
    After a write() to a regular file has successfully returned, any
    successful read() from each byte position on the file that was modified
    by that write() will return the data that was written by the write(). A
    subsequent write() to the same byte will overwrite the file data. If a
    read() of a file data can be proven by any means [e.g., MPI_Barrier()]
    to occur after a write() of that data, it must reflect that write(),
    even if the calls are made by a different process.
	Lewin, D. (1994). "POSIX Programmer's Guide (pg. 513-4)". O'Reilly
	& Associates.


Appendix A. Sample programs
---------------------------
Here are sample MPI-IO C and Fortran programs.  You may use them to run simple
tests of your MPI compilers and the parallel file system.  The MPI commands
used here are mpicc, mpif90 and mpiexec.  Replace them with the commands of
your system.

The programs assume they run in the parallel file system. Thus they create
the test data file in the current directory.  If the parallel file system
is somewhere else, you need to run the sample programs there or edit the
programs to use a different file name.

Example compiling and running:

% mpicc Sample_mpio.c -o c.out
% mpiexec -np 4 c.out

% mpif90 Sample_mpio.f90 -o f.out
% mpiexec -np 4 f.out


==> Sample_mpio.c <==
/* Simple MPI-IO program testing if a parallel file can be created.
 * Default filename can be specified via first program argument.
 * Each process writes something, then reads all data back.
 */

#include <mpi.h>
#ifndef MPI_FILE_NULL           /*MPIO may be defined in mpi.h already       */
#   include <mpio.h>
#endif

#define DIMSIZE	10		/* dimension size, avoid powers of 2. */
#define PRINTID printf("Proc %d: ", mpi_rank)

main(int ac, char **av)
{
    char hostname[128];
    int  mpi_size, mpi_rank;
    MPI_File fh;
    char *filename = "./mpitest.data";
    char mpi_err_str[MPI_MAX_ERROR_STRING];
    int  mpi_err_strlen;
    int  mpi_err;
    char writedata[DIMSIZE], readdata[DIMSIZE];
    char expect_val;
    int  i, irank; 
    int  nerrors = 0;		/* number of errors */
    MPI_Offset  mpi_off;
    MPI_Status  mpi_stat;

    MPI_Init(&ac, &av);
    MPI_Comm_size(MPI_COMM_WORLD, &mpi_size);
    MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);

    /* get file name if provided */
    if (ac > 1){
	filename = *++av;
    }
    if (mpi_rank==0){
	printf("Testing simple MPIO program with %d processes accessing file %s\n",
	    mpi_size, filename);
        printf("    (Filename can be specified via program argument)\n");
    }

    /* show the hostname so that we can tell where the processes are running */
    if (gethostname(hostname, 128) < 0){
	PRINTID;
	printf("gethostname failed\n");
	return 1;
    }
    PRINTID;
    printf("hostname=%s\n", hostname);

    if ((mpi_err = MPI_File_open(MPI_COMM_WORLD, filename,
	    MPI_MODE_RDWR | MPI_MODE_CREATE | MPI_MODE_DELETE_ON_CLOSE,
	    MPI_INFO_NULL, &fh))
	    != MPI_SUCCESS){
	MPI_Error_string(mpi_err, mpi_err_str, &mpi_err_strlen);
	PRINTID;
	printf("MPI_File_open failed (%s)\n", mpi_err_str);
	return 1;
    }

    /* each process writes some data */
    for (i=0; i < DIMSIZE; i++)
	writedata[i] = mpi_rank*DIMSIZE + i;
    mpi_off = mpi_rank*DIMSIZE;
    if ((mpi_err = MPI_File_write_at(fh, mpi_off, writedata, DIMSIZE, MPI_BYTE,
	    &mpi_stat))
	    != MPI_SUCCESS){
	MPI_Error_string(mpi_err, mpi_err_str, &mpi_err_strlen);
	PRINTID;
	printf("MPI_File_write_at offset(%ld), bytes (%d), failed (%s)\n",
		(long) mpi_off, (int) DIMSIZE, mpi_err_str);
	return 1;
    };

    /* make sure all processes has done writing. */
    MPI_Barrier(MPI_COMM_WORLD);

    /* each process reads all data and verify. */
    for (irank=0; irank < mpi_size; irank++){
	mpi_off = irank*DIMSIZE;
	if ((mpi_err = MPI_File_read_at(fh, mpi_off, readdata, DIMSIZE, MPI_BYTE,
		&mpi_stat))
		!= MPI_SUCCESS){
	    MPI_Error_string(mpi_err, mpi_err_str, &mpi_err_strlen);
	    PRINTID;
	    printf("MPI_File_read_at offset(%ld), bytes (%d), failed (%s)\n",
		    (long) mpi_off, (int) DIMSIZE, mpi_err_str);
	    return 1;
	};
	for (i=0; i < DIMSIZE; i++){
	    expect_val = irank*DIMSIZE + i;
	    if (readdata[i] != expect_val){
		PRINTID;
		printf("read data[%d:%d] got %d, expect %d\n", irank, i,
			readdata[i], expect_val);
		nerrors++;
	    }
	}
    }
    if (nerrors)
	return 1;

    MPI_File_close(&fh);

    PRINTID;
    printf("all tests passed\n");

    MPI_Finalize();
    return 0;
}

==> Sample_mpio.f90 <==
!
! The following example demonstrates how to create and close a parallel
! file using MPI-IO calls.
!
! USE MPI is the proper way to bring in MPI definitions but many
! MPI Fortran compiler supports the pseudo standard of INCLUDE.
! So, HDF5 uses the INCLUDE statement instead.
! 

     PROGRAM MPIOEXAMPLE

     USE mpi

     IMPLICIT NONE
        
     CHARACTER(LEN=80), PARAMETER :: filename = "filef.h5" ! File name
     INTEGER     ::   ierror  ! Error flag
     INTEGER     ::   fh      ! File handle
     INTEGER     ::   amode   ! File access mode
     
     call MPI_INIT(ierror)
     amode = MPI_MODE_RDWR + MPI_MODE_CREATE + MPI_MODE_DELETE_ON_CLOSE
     call MPI_FILE_OPEN(MPI_COMM_WORLD, filename, amode, MPI_INFO_NULL, fh, ierror)
     print *, "Trying to create ", filename
     if ( ierror .eq. MPI_SUCCESS ) then
        print *, "MPI_FILE_OPEN succeeded"
        call MPI_FILE_CLOSE(fh, ierror)
     else
        print *, "MPI_FILE_OPEN failed"
     endif

     call MPI_FINALIZE(ierror);
     END PROGRAM