Open MPI logo

Open MPI User's Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Open MPI User's mailing list

Subject: Re: [OMPI users] problem with groups and communicators in openmpi-1.6.4rc2
From: Ralph Castain (rhc_at_[hidden])
Date: 2013-01-19 10:28:36


Ah - cool! Thanks!

On Jan 19, 2013, at 7:19 AM, George Bosilca <bosilca_at_[hidden]> wrote:

> On Jan 19, 2013, at 15:44 , Ralph Castain <rhc_at_[hidden]> wrote:
>
>> I used your test code to confirm it also fails on our trunk - it looks like someone got the reference count wrong when creating/destructing groups.
>
> No, the code is not MPI compliant.
>
> The culprit is line 254 in the test code where Siegmar manually copied the group_comm_world into group_worker. This is correct as long as you remember that group_worker is not directly an MPI generated group, and as a result you are not allowed to free it.
>
> Now if you replace the:
>
> group_worker = group_comm_world
>
> by an MPI operation that create a copy of the original group such as
>
> MPI_Comm_group (MPI_COMM_WORLD, &group_worker);
>
> your code become MPI valid, and works without any issue in Open MPI.
>
> George.
>
>
>>
>> Afraid I'll have to defer to the authors of that code area...
>>
>>
>> On Jan 19, 2013, at 1:27 AM, Siegmar Gross <Siegmar.Gross_at_[hidden]> wrote:
>>
>>> Hi
>>>
>>> I have installed openmpi-1.6.4rc2 and have the following problem.
>>>
>>> tyr strided_vector 110 ompi_info | grep "Open MPI:"
>>> Open MPI: 1.6.4rc2r27861
>>> tyr strided_vector 111 mpicc -showme
>>> gcc -I/usr/local/openmpi-1.6.4_64_gcc/include -fexceptions -pthread -m64
>>> -L/usr/local/openmpi-1.6.4_64_gcc/lib64 -lmpi -lm -lkstat -llgrp -lsocket -lnsl
>>> -lrt -lm
>>>
>>>
>>> tyr strided_vector 112 mpiexec -np 4 data_type_4
>>> Process 2 of 4 running on tyr.informatik.hs-fulda.de
>>> Process 0 of 4 running on tyr.informatik.hs-fulda.de
>>> Process 3 of 4 running on tyr.informatik.hs-fulda.de
>>> Process 1 of 4 running on tyr.informatik.hs-fulda.de
>>>
>>> original matrix:
>>>
>>> 1 2 3 4 5 6 7 8 9 10
>>> 11 12 13 14 15 16 17 18 19 20
>>> 21 22 23 24 25 26 27 28 29 30
>>> 31 32 33 34 35 36 37 38 39 40
>>> 41 42 43 44 45 46 47 48 49 50
>>> 51 52 53 54 55 56 57 58 59 60
>>>
>>> result matrix:
>>> elements are sqared in columns:
>>> 0 1 2 6 7
>>> elements are multiplied with 2 in columns:
>>> 3 4 5 8 9
>>>
>>> 1 4 9 8 10 12 49 64 18 20
>>> 121 144 169 28 30 32 289 324 38 40
>>> 441 484 529 48 50 52 729 784 58 60
>>> 961 1024 1089 68 70 72 1369 1444 78 80
>>> 1681 1764 1849 88 90 92 2209 2304 98 100
>>> 2601 2704 2809 108 110 112 3249 3364 118 120
>>>
>>> Assertion failed: OPAL_OBJ_MAGIC_ID == ((opal_object_t *) (comm->c_remote_group)
>>> )->obj_magic_id, file ../../openmpi-1.6.4rc2r27861/ompi/communicator/comm_init.c
>>> , line 412
>>> [tyr:18578] *** Process received signal ***
>>> [tyr:18578] Signal: Abort (6)
>>> [tyr:18578] Signal code: (-1)
>>> Assertion failed: OPAL_OBJ_MAGIC_ID == ((opal_object_t *) (comm->c_remote_group)
>>> )->obj_magic_id, file ../../openmpi-1.6.4rc2r27861/ompi/communicator/comm_init.c
>>> , line 412
>>> /export2/prog/SunOS_sparc/openmpi-1.6.4_64_gcc/lib64/libmpi.so.1.0.7:opal_backtr
>>> ace_print+0x20
>>> [tyr:18580] *** Process received signal ***
>>> /export2/prog/SunOS_sparc/openmpi-1.6.4_64_gcc/lib64/libmpi.so.1.0.7:0x2c1bc4
>>> [tyr:18580] Signal: Abort (6)
>>> [tyr:18580] Signal code: (-1)
>>> /lib/sparcv9/libc.so.1:0xd88a4
>>> /lib/sparcv9/libc.so.1:0xcc418
>>> /lib/sparcv9/libc.so.1:0xcc624
>>> /lib/sparcv9/libc.so.1:__lwp_kill+0x8 [ Signal 6 (ABRT)]
>>> /lib/sparcv9/libc.so.1:abort+0xd0
>>> /lib/sparcv9/libc.so.1:_assert+0x74
>>> /export2/prog/SunOS_sparc/openmpi-1.6.4_64_gcc/lib64/libmpi.so.1.0.7:0xa4c58
>>> /export2/prog/SunOS_sparc/openmpi-1.6.4_64_gcc/lib64/libmpi.so.1.0.7:0xa2430
>>> /export2/prog/SunOS_sparc/openmpi-1.6.4_64_gcc/lib64/libmpi.so.1.0.7:ompi_comm_f
>>> inalize+0x168
>>> /export2/prog/SunOS_sparc/openmpi-1.6.4_64_gcc/lib64/libmpi.so.1.0.7:ompi_mpi_fi
>>> nalize+0xa60
>>> /export2/prog/SunOS_sparc/openmpi-1.6.4_64_gcc/lib64/libmpi.so.1.0.7:MPI_Finaliz
>>> e+0x90
>>> /home/fd1026/SunOS/sparc/bin/data_type_4:main+0x588
>>> /home/fd1026/SunOS/sparc/bin/data_type_4:_start+0x7c
>>> [tyr:18578] *** End of error message ***
>>> ...
>>>
>>>
>>>
>>> Everything works fine with LAM-MPI (even in a heterogeneous environment
>>> with little-endian and big-endian machines) so that it is probably an
>>> error in Open MPI (but you never know).
>>>
>>>
>>> tyr strided_vector 125 mpicc -showme
>>> gcc -I/usr/local/lam-6.5.9_64_gcc/include -L/usr/local/lam-6.5.9_64_gcc/lib
>>> -llamf77mpi -lmpi -llam -lsocket -lnsl
>>> tyr strided_vector 126 lamboot -v hosts.lam-mpi
>>>
>>> LAM 6.5.9/MPI 2 C++ - Indiana University
>>>
>>> Executing hboot on n0 (tyr.informatik.hs-fulda.de - 2 CPUs)...
>>> Executing hboot on n1 (sunpc1.informatik.hs-fulda.de - 4 CPUs)...
>>> topology done
>>>
>>> tyr strided_vector 127 mpirun -v app_data_type_4.lam-mpi
>>> 22894 data_type_4 running on local
>>> 22895 data_type_4 running on n0 (o)
>>> 21998 data_type_4 running on n1
>>> 22896 data_type_4 running on n0 (o)
>>> Process 1 of 4 running on tyr.informatik.hs-fulda.de
>>> Process 3 of 4 running on tyr.informatik.hs-fulda.de
>>> Process 2 of 4 running on sunpc1
>>> Process 0 of 4 running on tyr.informatik.hs-fulda.de
>>>
>>> original matrix:
>>>
>>> 1 2 3 4 5 6 7 8 9 10
>>> 11 12 13 14 15 16 17 18 19 20
>>> 21 22 23 24 25 26 27 28 29 30
>>> 31 32 33 34 35 36 37 38 39 40
>>> 41 42 43 44 45 46 47 48 49 50
>>> 51 52 53 54 55 56 57 58 59 60
>>>
>>> result matrix:
>>> elements are sqared in columns:
>>> 0 1 2 6 7
>>> elements are multiplied with 2 in columns:
>>> 3 4 5 8 9
>>>
>>> 1 4 9 8 10 12 49 64 18 20
>>> 121 144 169 28 30 32 289 324 38 40
>>> 441 484 529 48 50 52 729 784 58 60
>>> 961 1024 1089 68 70 72 1369 1444 78 80
>>> 1681 1764 1849 88 90 92 2209 2304 98 100
>>> 2601 2704 2809 108 110 112 3249 3364 118 120
>>>
>>> tyr strided_vector 128 lamhalt
>>>
>>> LAM 6.5.9/MPI 2 C++ - Indiana University
>>>
>>>
>>>
>>> I would be grateful, if somebody could fix the problem. Thank you
>>> very much for any help in advance.
>>>
>>>
>>> Kind regards
>>>
>>> Siegmar
>>> /* The program demonstrates how to set up and use a strided vector.
>>> * The process with rank 0 creates a matrix. The columns of the
>>> * matrix will then be distributed with a collective communication
>>> * operation to all processes. Each process performs an operation on
>>> * all column elements. Afterwards the results are collected in the
>>> * source matrix overwriting the original column elements.
>>> *
>>> * The program uses between one and n processes to change the values
>>> * of the column elements if the matrix has n columns. If you start
>>> * the program with one process it has to work on all n columns alone
>>> * and if you start it with n processes each process modifies the
>>> * values of one column. Every process must know how many columns it
>>> * has to modify so that it can allocate enough buffer space for its
>>> * column block. Therefore the process with rank 0 computes the
>>> * numbers of columns for each process in the array "num_columns" and
>>> * distributes this array with MPI_Broadcast to all processes. Each
>>> * process can now allocate memory for its column block. There is
>>> * still one task to do before the columns of the matrix can be
>>> * distributed with MPI_Scatterv: The size of every column block and
>>> * the offset of every column block must be computed und stored in
>>> * the arrays "sr_counts" and "sr_disps".
>>> *
>>> * An MPI data type is defined by its size, its contents, and its
>>> * extent. When multiple elements of the same size are used in a
>>> * contiguous manner (e.g. in a "scatter" operation or an operation
>>> * with "count" greater than one) the extent is used to compute where
>>> * the next element will start. The extent for a derived data type is
>>> * as big as the size of the derived data type so that the first
>>> * elements of the second structure will start after the last element
>>> * of the first structure, i.e., you have to "resize" the new data
>>> * type if you want to send it multiple times (count > 1) or to
>>> * scatter/gather it to many processes. Restrict the extent of the
>>> * derived data type for a strided vector in such a way that it looks
>>> * like just one element if it is used with "count > 1" or in a
>>> * scatter/gather operation.
>>> *
>>> * This version constructs a new column type (strided vector) with
>>> * "MPI_Type_vector" and uses collective communication. The new
>>> * data type knows the number of elements within one column and the
>>> * spacing between two column elements. The program uses at most
>>> * n processes if the matrix has n columns, i.e. depending on the
>>> * number of processes each process receives between 1 and n columns.
>>> * You can execute this program with an arbitrary number of processes
>>> * because it creates its own group with "num_worker" (<= n) processes
>>> * to perform the work if the matrix has n columns and the basic group
>>> * contains too many processes.
>>> *
>>> *
>>> * Compiling:
>>> * Store executable(s) into local directory.
>>> * mpicc -o <program name> <source code file name>
>>> *
>>> * Store executable(s) into predefined directories.
>>> * make
>>> *
>>> * Make program(s) automatically on all specified hosts. You must
>>> * edit the file "make_compile" and specify your host names before
>>> * you execute it.
>>> * make_compile
>>> *
>>> * Running:
>>> * LAM-MPI:
>>> * mpiexec -boot -np <number of processes> <program name>
>>> * or
>>> * mpiexec -boot \
>>> * -host <hostname> -np <number of processes> <program name> : \
>>> * -host <hostname> -np <number of processes> <program name>
>>> * or
>>> * mpiexec -boot [-v] -configfile <application file>
>>> * or
>>> * lamboot [-v] [<host file>]
>>> * mpiexec -np <number of processes> <program name>
>>> * or
>>> * mpiexec [-v] -configfile <application file>
>>> * lamhalt
>>> *
>>> * OpenMPI:
>>> * "host1", "host2", and so on can all have the same name,
>>> * if you want to start a virtual computer with some virtual
>>> * cpu's on the local host. The name "localhost" is allowed
>>> * as well.
>>> *
>>> * mpiexec -np <number of processes> <program name>
>>> * or
>>> * mpiexec --host <host1,host2,...> \
>>> * -np <number of processes> <program name>
>>> * or
>>> * mpiexec -hostfile <hostfile name> \
>>> * -np <number of processes> <program name>
>>> * or
>>> * mpiexec -app <application file>
>>> *
>>> * Cleaning:
>>> * local computer:
>>> * rm <program name>
>>> * or
>>> * make clean_all
>>> * on all specified computers (you must edit the file "make_clean_all"
>>> * and specify your host names before you execute it.
>>> * make_clean_all
>>> *
>>> *
>>> * File: data_type_4.c Author: S. Gross
>>> * Date: 30.08.2012
>>> *
>>> */
>>>
>>> #include <stdio.h>
>>> #include <stdlib.h>
>>> #include "mpi.h"
>>>
>>> #define P 6 /* # of rows */
>>> #define Q 10 /* # of columns */
>>> #define FACTOR 2 /* multiplicator for col. elem. */
>>> #define DEF_NUM_WORKER Q /* # of workers, must be <= Q */
>>>
>>> /* define macro to test the result of a "malloc" operation */
>>> #define TestEqualsNULL(val) \
>>> if (val == NULL) \
>>> { \
>>> fprintf (stderr, "file: %s line %d: Couldn't allocate memory.\n", \
>>> __FILE__, __LINE__); \
>>> exit (EXIT_FAILURE); \
>>> }
>>>
>>> /* define macro to determine the minimum of two values */
>>> #define MIN(a,b) ((a) < (b) ? (a) : (b))
>>>
>>>
>>> static void print_matrix (int p, int q, double **mat);
>>>
>>>
>>> int main (int argc, char *argv[])
>>> {
>>> int ntasks, /* number of parallel tasks */
>>> mytid, /* my task id */
>>> namelen, /* length of processor name */
>>> i, j, /* loop variables */
>>> *num_columns, /* # of columns in column block */
>>> *sr_counts, /* send/receive counts */
>>> *sr_disps, /* send/receive displacements */
>>> tmp, tmp1; /* temporary values */
>>> double matrix[P][Q],
>>> **col_block; /* column block of matrix */
>>> char processor_name[MPI_MAX_PROCESSOR_NAME];
>>> MPI_Datatype column_t, /* column type (strided vector) */
>>> col_block_t,
>>> tmp_column_t; /* needed to resize the extent */
>>> MPI_Group group_comm_world, /* processes in "basic group" */
>>> group_worker, /* processes in new groups */
>>> group_other;
>>> MPI_Comm COMM_WORKER, /* communicators for new groups */
>>> COMM_OTHER;
>>> int num_worker, /* # of worker in "group_worker"*/
>>> *group_w_mem, /* array of worker members */
>>> group_w_ntasks, /* # of tasks in "group_worker" */
>>> group_o_ntasks, /* # of tasks in "group_other" */
>>> group_w_mytid, /* my task id in "group_worker" */
>>> group_o_mytid, /* my task id in "group_other" */
>>> *universe_size_ptr, /* ptr to # of "virtual cpu's" */
>>> universe_size_flag; /* true if available */
>>>
>>> MPI_Init (&argc, &argv);
>>> MPI_Comm_rank (MPI_COMM_WORLD, &mytid);
>>> MPI_Comm_size (MPI_COMM_WORLD, &ntasks);
>>> /* Determine the correct number of processes for this program. If
>>> * there are more than Q processes (i.e., more processes than
>>> * columns) available, we split the "basic group" into two groups.
>>> * This program uses a group "group_worker" to do the real work
>>> * and a group "group_other" for the remaining processes of the
>>> * "basic group". The latter have nothing to do and can terminate
>>> * immediately. If there are less than or equal to Q processes
>>> * available all processes belong to group "group_worker" and group
>>> * "group_other" is empty. At first we find out which processes
>>> * belong to the "basic group".
>>> */
>>> MPI_Comm_group (MPI_COMM_WORLD, &group_comm_world);
>>> if (ntasks > Q)
>>> {
>>> /* There are too many processes, so that we must build a new group
>>> * with "num_worker" processes. "num_worker" will be the minimum of
>>> * DEF_NUM_WORKER and the "universe size" if it is supported by the
>>> * MPI implementation. At first we must check if DEF_NUM_WORKER has
>>> * a suitable value.
>>> */
>>> if (DEF_NUM_WORKER > Q)
>>> {
>>> if (mytid == 0)
>>> {
>>> fprintf (stderr, "\nError:\tInternal program error.\n"
>>> "\tConstant DEF_NUM_WORKER has value %d but must be\n"
>>> "\tlower than or equal to %d. Please change source\n"
>>> "\tcode and compile the program again.\n\n",
>>> DEF_NUM_WORKER, Q);
>>> }
>>> MPI_Group_free (&group_comm_world);
>>> MPI_Finalize ();
>>> exit (EXIT_FAILURE);
>>> }
>>> /* determine the universe size, set "num_worker" in an
>>> * appropriate way, and allocate memory for the array containing
>>> * the ranks of the members of the new group
>>> */
>>> MPI_Comm_get_attr (MPI_COMM_WORLD, MPI_UNIVERSE_SIZE,
>>> &universe_size_ptr, &universe_size_flag);
>>> if ((universe_size_flag != 0) && (*universe_size_ptr > 0))
>>> {
>>> num_worker = MIN (DEF_NUM_WORKER, *universe_size_ptr);
>>> }
>>> else
>>> {
>>> num_worker = DEF_NUM_WORKER;
>>> }
>>> group_w_mem = (int *) malloc (num_worker * sizeof (int));
>>> TestEqualsNULL (group_w_mem); /* test if memory was available */
>>> if (mytid == 0)
>>> {
>>> printf ("\nYou have started %d processes but I need at most "
>>> "%d processes.\n"
>>> "The universe contains %d \"virtual cpu's\" (\"0\" means "
>>> "not supported).\n"
>>> "I build a new worker group with %d processes. The "
>>> "processes with\n"
>>> "the following ranks in the basic group belong to "
>>> "the new group:\n ",
>>> ntasks, Q, *universe_size_ptr, num_worker);
>>> }
>>> for (i = 0; i < num_worker; ++i)
>>> {
>>> /* fetch some ranks from the basic group for the new worker
>>> * group, e.g. the last num_worker ranks to demonstrate that
>>> * a process may have different ranks in different groups
>>> */
>>> group_w_mem[i] = (ntasks - num_worker) + i;
>>> if (mytid == 0)
>>> {
>>> printf ("%d ", group_w_mem[i]);
>>> }
>>> }
>>> if (mytid == 0)
>>> {
>>> printf ("\n\n");
>>> }
>>> /* Create group "group_worker" */
>>> MPI_Group_incl (group_comm_world, num_worker, group_w_mem,
>>> &group_worker);
>>> free (group_w_mem);
>>> }
>>> else
>>> {
>>> /* there are at most as many processes as columns in our matrix,
>>> * i.e., we can use the "basic group"
>>> */
>>> group_worker = group_comm_world;
>>> }
>>> /* Create group "group_other" which demonstrates only how to use
>>> * another group operation and which has nothing to do in this
>>> * program.
>>> */
>>> MPI_Group_difference (group_comm_world, group_worker,
>>> &group_other);
>>> MPI_Group_free (&group_comm_world);
>>> /* Create communicators for both groups. The communicator is only
>>> * defined for all processes of the group and it is undefined
>>> * (MPI_COMM_NULL) for all other processes.
>>> */
>>> MPI_Comm_create (MPI_COMM_WORLD, group_worker, &COMM_WORKER);
>>> MPI_Comm_create (MPI_COMM_WORLD, group_other, &COMM_OTHER);
>>>
>>>
>>> /* =========================================================
>>> * ====== ======
>>> * ====== Supply work for all different groups. ======
>>> * ====== ======
>>> * ====== ======
>>> * ====== At first you must find out if a process ======
>>> * ====== belongs to a special group. You can use ======
>>> * ====== MPI_Group_rank for this purpose. It returns ======
>>> * ====== the rank of the calling process in the ======
>>> * ====== specified group or MPI_UNDEFINED if the ======
>>> * ====== calling process is not a member of the ======
>>> * ====== group. ======
>>> * ====== ======
>>> * =========================================================
>>> */
>>>
>>>
>>> /* =========================================================
>>> * ====== This is the group "group_worker". ======
>>> * =========================================================
>>> */
>>> MPI_Group_rank (group_worker, &group_w_mytid);
>>> if (group_w_mytid != MPI_UNDEFINED)
>>> {
>>> MPI_Comm_size (COMM_WORKER, &group_w_ntasks); /* # of processes */
>>> /* Now let's start with the real work */
>>> MPI_Get_processor_name (processor_name, &namelen);
>>> /* With the next statement every process executing this code will
>>> * print one line on the display. It may happen that the lines will
>>> * get mixed up because the display is a critical section. In general
>>> * only one process (mostly the process with rank 0) will print on
>>> * the display and all other processes will send their messages to
>>> * this process. Nevertheless for debugging purposes (or to
>>> * demonstrate that it is possible) it may be useful if every
>>> * process prints itself.
>>> */
>>> fprintf (stdout, "Process %d of %d running on %s\n",
>>> group_w_mytid, group_w_ntasks, processor_name);
>>> fflush (stdout);
>>> MPI_Barrier (COMM_WORKER); /* wait for all other processes */
>>>
>>> /* Build the new type for a strided vector and resize the extent
>>> * of the new datatype in such a way that the extent of the whole
>>> * column looks like just one element so that the next column
>>> * starts in matrix[0][i] in MPI_Scatterv/MPI_Gatherv.
>>> */
>>> MPI_Type_vector (P, 1, Q, MPI_DOUBLE, &tmp_column_t);
>>> MPI_Type_create_resized (tmp_column_t, 0, sizeof (double),
>>> &column_t);
>>> MPI_Type_commit (&column_t);
>>> MPI_Type_free (&tmp_column_t);
>>> if (group_w_mytid == 0)
>>> {
>>> tmp = 1;
>>> for (i = 0; i < P; ++i) /* initialize matrix */
>>> {
>>> for (j = 0; j < Q; ++j)
>>> {
>>> matrix[i][j] = tmp++;
>>> }
>>> }
>>> printf ("\n\noriginal matrix:\n\n");
>>> print_matrix (P, Q, (double **) matrix);
>>> }
>>> /* allocate memory for array containing the number of columns of a
>>> * column block for each process
>>> */
>>> num_columns = (int *) malloc (group_w_ntasks * sizeof (int));
>>> TestEqualsNULL (num_columns); /* test if memory was available */
>>>
>>> /* do an unnecessary initialization to make the GNU compiler happy
>>> * so that you won't get a warning about the use of a possibly
>>> * uninitialized variable
>>> */
>>> sr_counts = NULL;
>>> sr_disps = NULL;
>>> if (group_w_mytid == 0)
>>> {
>>> /* allocate memory for arrays containing the size and
>>> * displacement of each column block
>>> */
>>> sr_counts = (int *) malloc (group_w_ntasks * sizeof (int));
>>> TestEqualsNULL (sr_counts);
>>> sr_disps = (int *) malloc (group_w_ntasks * sizeof (int));
>>> TestEqualsNULL (sr_disps);
>>> /* compute number of columns in column block for each process */
>>> tmp = Q / group_w_ntasks;
>>> for (i = 0; i < group_w_ntasks; ++i)
>>> {
>>> num_columns[i] = tmp; /* number of columns */
>>> }
>>> for (i = 0; i < (Q % group_w_ntasks); ++i) /* adjust size */
>>> {
>>> num_columns[i]++;
>>> }
>>> for (i = 0; i < group_w_ntasks; ++i)
>>> {
>>> /* nothing to do because "column_t" contains already all
>>> * elements of a column, i.e., the "size" is equal to the
>>> * number of columns in the block
>>> */
>>> sr_counts[i] = num_columns[i]; /* "size" of column-block */
>>> }
>>> sr_disps[0] = 0; /* start of i-th column-block */
>>> for (i = 1; i < group_w_ntasks; ++i)
>>> {
>>> sr_disps[i] = sr_disps[i - 1] + sr_counts[i - 1];
>>> }
>>> }
>>> /* inform all processes about their column block sizes */
>>> MPI_Bcast (num_columns, group_w_ntasks, MPI_INT, 0, COMM_WORKER);
>>> /* allocate memory for a column block and define a new derived
>>> * data type for the column block. This data type is possibly
>>> * different for different processes if the number of processes
>>> * isn't a factor of the row size of the original matrix. Don't
>>> * forget to resize the extent of the new data type in such a
>>> * way that the extent of the whole column looks like just one
>>> * element so that the next column starts in col_block[0][i]
>>> * in MPI_Scatterv/MPI_Gatherv.
>>> */
>>> col_block = (double **) malloc (P * num_columns[group_w_mytid] *
>>> sizeof (double));
>>> TestEqualsNULL (col_block);
>>> MPI_Type_vector (P, 1, num_columns[group_w_mytid], MPI_DOUBLE,
>>> &tmp_column_t);
>>> MPI_Type_create_resized (tmp_column_t, 0, sizeof (double),
>>> &col_block_t);
>>> MPI_Type_commit (&col_block_t);
>>> MPI_Type_free (&tmp_column_t);
>>> /* send column block i of "matrix" to process i */
>>> MPI_Scatterv (matrix, sr_counts, sr_disps, column_t,
>>> col_block, num_columns[group_w_mytid],
>>> col_block_t, 0, COMM_WORKER);
>>> /* Modify column elements. The compiler doesn't know the structure
>>> * of the column block matrix so that you have to do the index
>>> * calculations for mat[i][j] yourself. In C a matrix is stored
>>> * row-by-row so that the i-th row starts at location "i * q" if
>>> * the matrix has "q" columns. Therefore the address of mat[i][j]
>>> * can be expressed as "(double *) mat + i * q + j" and mat[i][j]
>>> * itself as "*((double *) mat + i * q + j)".
>>> */
>>> for (i = 0; i < P; ++i)
>>> {
>>> for (j = 0; j < num_columns[group_w_mytid]; ++j)
>>> {
>>> if ((group_w_mytid % 2) == 0)
>>> {
>>> /* col_block[i][j] *= col_block[i][j] */
>>>
>>> *((double *) col_block + i * num_columns[group_w_mytid] + j) *=
>>> *((double *) col_block + i * num_columns[group_w_mytid] + j);
>>> }
>>> else
>>> {
>>> /* col_block[i][j] *= FACTOR */
>>>
>>> *((double *) col_block + i * num_columns[group_w_mytid] + j) *=
>>> FACTOR;
>>> }
>>> }
>>> }
>>> /* receive column-block i of "matrix" from process i */
>>> MPI_Gatherv (col_block, num_columns[group_w_mytid], col_block_t,
>>> matrix, sr_counts, sr_disps, column_t,
>>> 0, COMM_WORKER);
>>> if (group_w_mytid == 0)
>>> {
>>> printf ("\n\nresult matrix:\n"
>>> " elements are sqared in columns:\n ");
>>> tmp = 0;
>>> tmp1 = 0;
>>> for (i = 0; i < group_w_ntasks; ++i)
>>> {
>>> tmp1 = tmp1 + num_columns[i];
>>> if ((i % 2) == 0)
>>> {
>>> for (j = tmp; j < tmp1; ++j)
>>> {
>>> printf ("%4d", j);
>>> }
>>> }
>>> tmp = tmp1;
>>> }
>>> printf ("\n elements are multiplied with %d in columns:\n ",
>>> FACTOR);
>>> tmp = 0;
>>> tmp1 = 0;
>>> for (i = 0; i < group_w_ntasks; ++i)
>>> {
>>> tmp1 = tmp1 + num_columns[i];
>>> if ((i % 2) != 0)
>>> {
>>> for (j = tmp; j < tmp1; ++j)
>>> {
>>> printf ("%4d", j);
>>> }
>>> }
>>> tmp = tmp1;
>>> }
>>> printf ("\n\n\n");
>>> print_matrix (P, Q, (double **) matrix);
>>> free (sr_counts);
>>> free (sr_disps);
>>> }
>>> free (num_columns);
>>> free (col_block);
>>> MPI_Type_free (&column_t);
>>> MPI_Type_free (&col_block_t);
>>> MPI_Comm_free (&COMM_WORKER);
>>> }
>>>
>>>
>>> /* =========================================================
>>> * ====== This is the group "group_other". ======
>>> * =========================================================
>>> */
>>> MPI_Group_rank (group_other, &group_o_mytid);
>>> if (group_o_mytid != MPI_UNDEFINED)
>>> {
>>> /* Nothing to do (only to demonstrate how to divide work for
>>> * different groups).
>>> */
>>> MPI_Comm_size (COMM_OTHER, &group_o_ntasks);
>>> if (group_o_mytid == 0)
>>> {
>>> if (group_o_ntasks == 1)
>>> {
>>> printf ("\nGroup \"group_other\" contains %d process "
>>> "which has\n"
>>> "nothing to do.\n\n", group_o_ntasks);
>>> }
>>> else
>>> {
>>> printf ("\nGroup \"group_other\" contains %d processes "
>>> "which have\n"
>>> "nothing to do.\n\n", group_o_ntasks);
>>> }
>>> }
>>> MPI_Comm_free (&COMM_OTHER);
>>> }
>>>
>>>
>>> /* =========================================================
>>> * ====== all groups will reach this point ======
>>> * =========================================================
>>> */
>>> MPI_Group_free (&group_worker);
>>> MPI_Group_free (&group_other);
>>> MPI_Finalize ();
>>> return EXIT_SUCCESS;
>>> }
>>>
>>>
>>> /* Print the values of an arbitrary 2D-matrix of "double" values. The
>>> * compiler doesn't know the structure of the matrix so that you have
>>> * to do the index calculations for mat[i][j] yourself. In C a matrix
>>> * is stored row-by-row so that the i-th row starts at location "i * q"
>>> * if the matrix has "q" columns. Therefore the address of mat[i][j]
>>> * can be expressed as "(double *) mat + i * q + j" and mat[i][j]
>>> * itself as "*((double *) mat + i * q + j)".
>>> *
>>> * input parameters: p number of rows
>>> * q number of columns
>>> * mat 2D-matrix of "double" values
>>> * output parameters: none
>>> * return value: none
>>> * side effects: none
>>> *
>>> */
>>> void print_matrix (int p, int q, double **mat)
>>> {
>>> int i, j; /* loop variables */
>>>
>>> for (i = 0; i < p; ++i)
>>> {
>>> for (j = 0; j < q; ++j)
>>> {
>>> printf ("%6g", *((double *) mat + i * q + j));
>>> }
>>> printf ("\n");
>>> }
>>> printf ("\n");
>>> }
>>> _______________________________________________
>>> users mailing list
>>> users_at_[hidden]
>>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>>
>>
>> _______________________________________________
>> users mailing list
>> users_at_[hidden]
>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>
>
> _______________________________________________
> users mailing list
> users_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/users