April 2014 - discuss - mailman3.cels.anl.gov

Re: [mpich-discuss] MPI-IO bug
by Rob Latham 30 Apr '14

30 Apr '14

Wei-keng, I've already let this slip longer than I intended. I've opened a ticket [1], which I intend to close pretty quickly http://trac.mpich.org/projects/mpich/ticket/2073 ==rob On 04/09/2014 02:52 PM, Wei-keng Liao wrote: > (This bug is probably caused by my patch long ago.) > Attached is a program extracted from an application that can reproduce > the problem > observed from a large run. The problem is when defining a filetype using > MPI_Type_indexed > and the first few elements of argument blocklens[] are zeros, a > collective write will > miss writing some data. > > The test program first fills a file with 9 integers with values all -999. > It then defines a filetype and writes to the file in parallel with user > buffers > with value all 1s. Lastly, the file is read back and checked for contents. > > Command used to compile and run the test program: > mpicc -g -o bug_indexed_io bug_indexed_io.c > mpiexec -n 2 bug_indexed_io > > Stdout: > 0: Error: unexpected varlue at buf[7] == -999 > > > The patch below can fix this problem. Hope it does not break other tests. > > Index: adio/common/ad_read_coll.c > @@ -368,13 +368,16 @@ > #endif > if (file_ptr_type == ADIO_INDIVIDUAL) { > /* Wei-keng reworked type processing to be a bit more > efficient */ > + for (i=0; i<flat_file->count; i++) /* skip blocklens[] == 0 */ > + if (flat_file->blocklens[i] > 0) break; > + > offset = fd->fp_ind - disp; > - n_filetypes = (offset - flat_file->indices[0]) / > filetype_extent; > + n_filetypes = (offset - flat_file->indices[i]) / > filetype_extent; > offset -= (ADIO_Offset)n_filetypes * filetype_extent; > /* now offset is local to this extent */ > > /* find the block where offset is located, skip > blocklens[i]==0 */ > - for (i=0; i<flat_file->count; i++) { > + for (; i<flat_file->count; i++) { > ADIO_Offset dist; > if (flat_file->blocklens[i] == 0) continue; > dist = flat_file->indices[i] + flat_file->blocklens[i] > - offset; > > Wei-keng > > > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss -- Rob Latham Mathematics and Computer Science Division Argonne National Lab, IL USA

1 0

Re: [mpich-discuss] Memory leak comparing communicators
by VandeVondele Joost 28 Apr '14

28 Apr '14

I'm not quite sure, not for the bulk of the work, but maybe in its parallel libraries (e.g. scalapack). >From the trace, at least 'if (comm_ptr1->comm_kind == MPID_INTRACOMM) ' seems true. It does use cartesian communicators in these compares, but this didn't trigger it in the testcase. Joost On 04/28/2014 02:34 PM, VandeVondele Joost wrote: > Unfortunately, the leak doesn't reproduce with some testcases I tried to construct. I'm not sure what makes the real application different from the testcases, so I don't have much more info than the backtraces. > is cp2k using intercommunicators? ==rob -- Rob Latham Mathematics and Computer Science Division Argonne National Lab, IL USA _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] Memory leak comparing communicators
by Rob Latham 28 Apr '14

28 Apr '14

On 04/28/2014 02:34 PM, VandeVondele Joost wrote: > Unfortunately, the leak doesn't reproduce with some testcases I tried to construct. I'm not sure what makes the real application different from the testcases, so I don't have much more info than the backtraces. > is cp2k using intercommunicators? ==rob -- Rob Latham Mathematics and Computer Science Division Argonne National Lab, IL USA

1 0

Re: [mpich-discuss] Memory leak comparing communicators
by VandeVondele Joost 28 Apr '14

28 Apr '14

Unfortunately, the leak doesn't reproduce with some testcases I tried to construct. I'm not sure what makes the real application different from the testcases, so I don't have much more info than the backtraces. ________________________________________ From: discuss-bounces(a)mpich.org [discuss-bounces(a)mpich.org] on behalf of VandeVondele Joost [joost.vandevondele(a)mat.ethz.ch] Sent: Monday, April 28, 2014 6:33 PM To: discuss(a)mpich.org Subject: Re: [mpich-discuss] Memory leak comparing communicators Hi Antonio, right now, a runnable example is or full CP2K program, freely available at cp2k.org, but not very practical as a testcase. It seems to happen at any point where we call: CALL mpi_comm_compare ( comm1, comm2, iout, ierr ) I'll see if I can reproduce it running something from the test directory. Joost ________________________________________ From: discuss-bounces(a)mpich.org [discuss-bounces(a)mpich.org] on behalf of "Antonio J. Peña" [apenya(a)mcs.anl.gov] Sent: Monday, April 28, 2014 6:26 PM To: discuss(a)mpich.org Subject: Re: [mpich-discuss] Memory leak comparing communicators Hi Joost, Can you share the smallest fragment of code with which we can reproduce this? Antonio On 04/28/2014 11:25 AM, VandeVondele Joost wrote: > Hi, > > using mpich 3.1 I notice that my application leaks memory. Compiling it with '-O1 -g -fno-omit-frame-pointer -fsanitize=leak' using gcc 4.9 it appears the sources is comparing communicators, for example: > > [0] Direct leak of 38400 byte(s) in 600 object(s) allocated from: > [0] #0 0x7f84298d53a8 in __interceptor_malloc ../../../../gcc/libsanitizer/lsan/lsan_interceptors.cc:66 > [0] #1 0x7f842aabaf6b in MPIR_Group_create /data/vjoost/mpich-3.1/src/mpi/group/grouputil.c:77 > [0] #2 0x7f842ab88314 in MPIR_Comm_group_impl /data/vjoost/mpich-3.1/src/mpi/comm/comm_group.c:44 > [0] #3 0x7f842ab840d3 in PMPI_Comm_compare /data/vjoost/mpich-3.1/src/mpi/comm/comm_compare.c:122 > [0] #4 0x7f842aac8e84 in pmpi_comm_compare_ /data/vjoost/mpich-3.1/src/binding/f77/comm_comparef.c:267 > [0] #5 0x1bfd447 in __message_passing_MOD_mp_comm_compare /data/vjoost/clean/cp2k/cp2k/src/common/message_passing.F:1084 > > I'm not seeing other leaks, so is guess something particular is going on here. Any suggestions on a possible fix ? > > Thanks, > > Joost > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss -- Antonio J. Peña Postdoctoral Appointee Mathematics and Computer Science Division Argonne National Laboratory 9700 South Cass Avenue, Bldg. 240, Of. 3148 Argonne, IL 60439-4847 apenya(a)mcs.anl.gov www.mcs.anl.gov/~apenya _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] mpi windows with groups mpich2
by Helen Kershaw 28 Apr '14

28 Apr '14

I've just built 3.1 on my laptop and the program runs fine. I'll have to bug the sys admins to upgrade to MPICH3.1 on the machine I want to run on. Thanks for you time, Helen On 4/28/14 11:27 AM, Helen Kershaw wrote: > Will do. > > On 4/28/14 11:24 AM, Rajeev Thakur wrote: >> Can you try the latest, 3.1. >> >> Rajeev >> >> On Apr 28, 2014, at 12:21 PM, Helen Kershaw <hkershaw(a)ucar.edu> >> wrote: >> >>> Version 3.0.4. >>> >>> Helen >>> >>> On 4/28/14 10:38 AM, Rajeev Thakur wrote: >>>> The program runs fine on my Mac laptop. Which version of MPICH are >>>> you using? >>>> >>>> Rajeev >>>> >>>> On Apr 28, 2014, at 10:41 AM, Helen Kershaw <hkershaw(a)ucar.edu> wrote: >>>> >>>>> Hi, >>>>> >>>>> I'm trying to use MPI one-sided communication with sub-communicators. >>>>> I've put my test code at the end of the email. It runs on 8 >>>>> processors, creates two communicators, makes a window on the >>>>> communicators, then tries to do some one-sided communication. >>>>> >>>>> I can use groups OR windows without a problem, but when I use them >>>>> together (in the test code) I get a coredump from the first 4 >>>>> tasks. I >>>>> think the one-sided communication may be using the local rank in >>>>> the sub-communicator (mpi_comm_grid) as if it is the global rank in >>>>> mpi_comm_world. >>>>> >>>>> I can run the test code ok using open-mpi, but not mpich2. I might be >>>>> missing something obvious. >>>>> >>>>> Thanks for your help, >>>>> Helen >>>>> >>>>> ----- >>>>> program test_group_win >>>>> >>>>> use mpi >>>>> >>>>> implicit none >>>>> >>>>> ! mpi >>>>> integer :: rank, size, ierr, status(mpi_status_size) >>>>> ! groups >>>>> integer, parameter :: group_size = 4 >>>>> integer :: group_members(group_size) >>>>> integer :: group_all ! mpi_comm_world group >>>>> integer :: subgroup >>>>> integer :: mpi_comm_grid >>>>> integer :: local_rank ! rank within subgroup >>>>> >>>>> >>>>> ! window >>>>> integer :: win >>>>> integer ::bytesize !> size in bytes of each >>>>> element in the window >>>>> integer(KIND=MPI_ADDRESS_KIND) :: window_size >>>>> pointer(aa, duplicate) >>>>> real :: duplicate(*) >>>>> integer(KIND=MPI_ADDRESS_KIND) :: target_disp !> displacement in >>>>> window >>>>> integer :: owner >>>>> real :: result >>>>> >>>>> call mpi_init(ierr) >>>>> call mpi_comm_size(mpi_comm_world, size, ierr) >>>>> call mpi_comm_rank(mpi_comm_world, rank, ierr) >>>>> >>>>> ! create groups >>>>> call mpi_comm_group(mpi_comm_world, group_all, ierr) >>>>> >>>>> if (rank < 4 ) then >>>>> group_members = (/0,1,2,3/) >>>>> else >>>>> group_members = (/4,5,6,7/) >>>>> endif >>>>> >>>>> call mpi_group_incl(group_all, group_size, group_members, subgroup, >>>>> ierr) >>>>> call mpi_comm_create(mpi_comm_world, subgroup, mpi_comm_grid, ierr) >>>>> call mpi_comm_rank(mpi_comm_grid, local_rank, ierr) ! rank within >>>>> group >>>>> >>>>> ! create window >>>>> call mpi_type_size(mpi_real, bytesize, ierr) >>>>> window_size = bytesize ! one element in the window >>>>> aa = malloc(1) >>>>> call MPI_ALLOC_MEM(window_size, mpi_info_null, aa, ierr) >>>>> >>>>> duplicate(1) = rank >>>>> >>>>> call mpi_win_create(duplicate, window_size, bytesize, MPI_INFO_NULL, >>>>> mpi_comm_grid, win, ierr) >>>>> >>>>> ! grabbing data from rank above >>>>> if (rank < 3 ) then >>>>> owner = rank + 1 >>>>> elseif (rank < 7) then >>>>> owner = rank + 1 - 4 >>>>> else >>>>> owner = 0 >>>>> endif >>>>> >>>>> print*, 'rank owner', rank, owner >>>>> >>>>> target_disp = 0 >>>>> >>>>> call mpi_win_lock(MPI_LOCK_SHARED, owner, 0, win, ierr) >>>>> call mpi_get(result, 1, mpi_real, owner, target_disp, 1, mpi_real, >>>>> win, >>>>> ierr) >>>>> call mpi_win_unlock(owner, win, ierr) >>>>> >>>>> print*, 'rank, result ', rank, result >>>>> >>>>> ! free window >>>>> call mpi_win_free(win, ierr) >>>>> call mpi_free_mem(duplicate, ierr) >>>>> >>>>> call mpi_finalize(ierr) >>>>> >>>>> end program test_group_win >>>>> ---- >>>>> >>>>> >>>>> _______________________________________________ >>>>> discuss mailing list discuss(a)mpich.org >>>>> To manage subscription options or unsubscribe: >>>>> https://lists.mpich.org/mailman/listinfo/discuss >>>> >>>> _______________________________________________ >>>> discuss mailing list discuss(a)mpich.org >>>> To manage subscription options or unsubscribe: >>>> https://lists.mpich.org/mailman/listinfo/discuss >>>> >>> _______________________________________________ >>> discuss mailing list discuss(a)mpich.org >>> To manage subscription options or unsubscribe: >>> https://lists.mpich.org/mailman/listinfo/discuss >> >> _______________________________________________ >> discuss mailing list discuss(a)mpich.org >> To manage subscription options or unsubscribe: >> https://lists.mpich.org/mailman/listinfo/discuss >> > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] mpi windows with groups mpich2
by Helen Kershaw 28 Apr '14

28 Apr '14

Will do. On 4/28/14 11:24 AM, Rajeev Thakur wrote: > Can you try the latest, 3.1. > > Rajeev > > On Apr 28, 2014, at 12:21 PM, Helen Kershaw <hkershaw(a)ucar.edu> > wrote: > >> Version 3.0.4. >> >> Helen >> >> On 4/28/14 10:38 AM, Rajeev Thakur wrote: >>> The program runs fine on my Mac laptop. Which version of MPICH are you using? >>> >>> Rajeev >>> >>> On Apr 28, 2014, at 10:41 AM, Helen Kershaw <hkershaw(a)ucar.edu> wrote: >>> >>>> Hi, >>>> >>>> I'm trying to use MPI one-sided communication with sub-communicators. >>>> I've put my test code at the end of the email. It runs on 8 processors, creates two communicators, makes a window on the communicators, then tries to do some one-sided communication. >>>> >>>> I can use groups OR windows without a problem, but when I use them >>>> together (in the test code) I get a coredump from the first 4 tasks. I >>>> think the one-sided communication may be using the local rank in the sub-communicator (mpi_comm_grid) as if it is the global rank in mpi_comm_world. >>>> >>>> I can run the test code ok using open-mpi, but not mpich2. I might be >>>> missing something obvious. >>>> >>>> Thanks for your help, >>>> Helen >>>> >>>> ----- >>>> program test_group_win >>>> >>>> use mpi >>>> >>>> implicit none >>>> >>>> ! mpi >>>> integer :: rank, size, ierr, status(mpi_status_size) >>>> ! groups >>>> integer, parameter :: group_size = 4 >>>> integer :: group_members(group_size) >>>> integer :: group_all ! mpi_comm_world group >>>> integer :: subgroup >>>> integer :: mpi_comm_grid >>>> integer :: local_rank ! rank within subgroup >>>> >>>> >>>> ! window >>>> integer :: win >>>> integer ::bytesize !> size in bytes of each >>>> element in the window >>>> integer(KIND=MPI_ADDRESS_KIND) :: window_size >>>> pointer(aa, duplicate) >>>> real :: duplicate(*) >>>> integer(KIND=MPI_ADDRESS_KIND) :: target_disp !> displacement in window >>>> integer :: owner >>>> real :: result >>>> >>>> call mpi_init(ierr) >>>> call mpi_comm_size(mpi_comm_world, size, ierr) >>>> call mpi_comm_rank(mpi_comm_world, rank, ierr) >>>> >>>> ! create groups >>>> call mpi_comm_group(mpi_comm_world, group_all, ierr) >>>> >>>> if (rank < 4 ) then >>>> group_members = (/0,1,2,3/) >>>> else >>>> group_members = (/4,5,6,7/) >>>> endif >>>> >>>> call mpi_group_incl(group_all, group_size, group_members, subgroup, ierr) >>>> call mpi_comm_create(mpi_comm_world, subgroup, mpi_comm_grid, ierr) >>>> call mpi_comm_rank(mpi_comm_grid, local_rank, ierr) ! rank within group >>>> >>>> ! create window >>>> call mpi_type_size(mpi_real, bytesize, ierr) >>>> window_size = bytesize ! one element in the window >>>> aa = malloc(1) >>>> call MPI_ALLOC_MEM(window_size, mpi_info_null, aa, ierr) >>>> >>>> duplicate(1) = rank >>>> >>>> call mpi_win_create(duplicate, window_size, bytesize, MPI_INFO_NULL, >>>> mpi_comm_grid, win, ierr) >>>> >>>> ! grabbing data from rank above >>>> if (rank < 3 ) then >>>> owner = rank + 1 >>>> elseif (rank < 7) then >>>> owner = rank + 1 - 4 >>>> else >>>> owner = 0 >>>> endif >>>> >>>> print*, 'rank owner', rank, owner >>>> >>>> target_disp = 0 >>>> >>>> call mpi_win_lock(MPI_LOCK_SHARED, owner, 0, win, ierr) >>>> call mpi_get(result, 1, mpi_real, owner, target_disp, 1, mpi_real, win, >>>> ierr) >>>> call mpi_win_unlock(owner, win, ierr) >>>> >>>> print*, 'rank, result ', rank, result >>>> >>>> ! free window >>>> call mpi_win_free(win, ierr) >>>> call mpi_free_mem(duplicate, ierr) >>>> >>>> call mpi_finalize(ierr) >>>> >>>> end program test_group_win >>>> ---- >>>> >>>> >>>> _______________________________________________ >>>> discuss mailing list discuss(a)mpich.org >>>> To manage subscription options or unsubscribe: >>>> https://lists.mpich.org/mailman/listinfo/discuss >>> >>> _______________________________________________ >>> discuss mailing list discuss(a)mpich.org >>> To manage subscription options or unsubscribe: >>> https://lists.mpich.org/mailman/listinfo/discuss >>> >> _______________________________________________ >> discuss mailing list discuss(a)mpich.org >> To manage subscription options or unsubscribe: >> https://lists.mpich.org/mailman/listinfo/discuss > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss >

1 0

Re: [mpich-discuss] mpi windows with groups mpich2
by Rajeev Thakur 28 Apr '14

28 Apr '14

Can you try the latest, 3.1. Rajeev On Apr 28, 2014, at 12:21 PM, Helen Kershaw <hkershaw(a)ucar.edu> wrote: > Version 3.0.4. > > Helen > > On 4/28/14 10:38 AM, Rajeev Thakur wrote: >> The program runs fine on my Mac laptop. Which version of MPICH are you using? >> >> Rajeev >> >> On Apr 28, 2014, at 10:41 AM, Helen Kershaw <hkershaw(a)ucar.edu> wrote: >> >>> Hi, >>> >>> I'm trying to use MPI one-sided communication with sub-communicators. >>> I've put my test code at the end of the email. It runs on 8 processors, creates two communicators, makes a window on the communicators, then tries to do some one-sided communication. >>> >>> I can use groups OR windows without a problem, but when I use them >>> together (in the test code) I get a coredump from the first 4 tasks. I >>> think the one-sided communication may be using the local rank in the sub-communicator (mpi_comm_grid) as if it is the global rank in mpi_comm_world. >>> >>> I can run the test code ok using open-mpi, but not mpich2. I might be >>> missing something obvious. >>> >>> Thanks for your help, >>> Helen >>> >>> ----- >>> program test_group_win >>> >>> use mpi >>> >>> implicit none >>> >>> ! mpi >>> integer :: rank, size, ierr, status(mpi_status_size) >>> ! groups >>> integer, parameter :: group_size = 4 >>> integer :: group_members(group_size) >>> integer :: group_all ! mpi_comm_world group >>> integer :: subgroup >>> integer :: mpi_comm_grid >>> integer :: local_rank ! rank within subgroup >>> >>> >>> ! window >>> integer :: win >>> integer ::bytesize !> size in bytes of each >>> element in the window >>> integer(KIND=MPI_ADDRESS_KIND) :: window_size >>> pointer(aa, duplicate) >>> real :: duplicate(*) >>> integer(KIND=MPI_ADDRESS_KIND) :: target_disp !> displacement in window >>> integer :: owner >>> real :: result >>> >>> call mpi_init(ierr) >>> call mpi_comm_size(mpi_comm_world, size, ierr) >>> call mpi_comm_rank(mpi_comm_world, rank, ierr) >>> >>> ! create groups >>> call mpi_comm_group(mpi_comm_world, group_all, ierr) >>> >>> if (rank < 4 ) then >>> group_members = (/0,1,2,3/) >>> else >>> group_members = (/4,5,6,7/) >>> endif >>> >>> call mpi_group_incl(group_all, group_size, group_members, subgroup, ierr) >>> call mpi_comm_create(mpi_comm_world, subgroup, mpi_comm_grid, ierr) >>> call mpi_comm_rank(mpi_comm_grid, local_rank, ierr) ! rank within group >>> >>> ! create window >>> call mpi_type_size(mpi_real, bytesize, ierr) >>> window_size = bytesize ! one element in the window >>> aa = malloc(1) >>> call MPI_ALLOC_MEM(window_size, mpi_info_null, aa, ierr) >>> >>> duplicate(1) = rank >>> >>> call mpi_win_create(duplicate, window_size, bytesize, MPI_INFO_NULL, >>> mpi_comm_grid, win, ierr) >>> >>> ! grabbing data from rank above >>> if (rank < 3 ) then >>> owner = rank + 1 >>> elseif (rank < 7) then >>> owner = rank + 1 - 4 >>> else >>> owner = 0 >>> endif >>> >>> print*, 'rank owner', rank, owner >>> >>> target_disp = 0 >>> >>> call mpi_win_lock(MPI_LOCK_SHARED, owner, 0, win, ierr) >>> call mpi_get(result, 1, mpi_real, owner, target_disp, 1, mpi_real, win, >>> ierr) >>> call mpi_win_unlock(owner, win, ierr) >>> >>> print*, 'rank, result ', rank, result >>> >>> ! free window >>> call mpi_win_free(win, ierr) >>> call mpi_free_mem(duplicate, ierr) >>> >>> call mpi_finalize(ierr) >>> >>> end program test_group_win >>> ---- >>> >>> >>> _______________________________________________ >>> discuss mailing list discuss(a)mpich.org >>> To manage subscription options or unsubscribe: >>> https://lists.mpich.org/mailman/listinfo/discuss >> >> _______________________________________________ >> discuss mailing list discuss(a)mpich.org >> To manage subscription options or unsubscribe: >> https://lists.mpich.org/mailman/listinfo/discuss >> > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] mpi windows with groups mpich2
by Helen Kershaw 28 Apr '14

28 Apr '14

Version 3.0.4. Helen On 4/28/14 10:38 AM, Rajeev Thakur wrote: > The program runs fine on my Mac laptop. Which version of MPICH are you using? > > Rajeev > > On Apr 28, 2014, at 10:41 AM, Helen Kershaw <hkershaw(a)ucar.edu> wrote: > >> Hi, >> >> I'm trying to use MPI one-sided communication with sub-communicators. >> I've put my test code at the end of the email. It runs on 8 processors, creates two communicators, makes a window on the communicators, then tries to do some one-sided communication. >> >> I can use groups OR windows without a problem, but when I use them >> together (in the test code) I get a coredump from the first 4 tasks. I >> think the one-sided communication may be using the local rank in the sub-communicator (mpi_comm_grid) as if it is the global rank in mpi_comm_world. >> >> I can run the test code ok using open-mpi, but not mpich2. I might be >> missing something obvious. >> >> Thanks for your help, >> Helen >> >> ----- >> program test_group_win >> >> use mpi >> >> implicit none >> >> ! mpi >> integer :: rank, size, ierr, status(mpi_status_size) >> ! groups >> integer, parameter :: group_size = 4 >> integer :: group_members(group_size) >> integer :: group_all ! mpi_comm_world group >> integer :: subgroup >> integer :: mpi_comm_grid >> integer :: local_rank ! rank within subgroup >> >> >> ! window >> integer :: win >> integer ::bytesize !> size in bytes of each >> element in the window >> integer(KIND=MPI_ADDRESS_KIND) :: window_size >> pointer(aa, duplicate) >> real :: duplicate(*) >> integer(KIND=MPI_ADDRESS_KIND) :: target_disp !> displacement in window >> integer :: owner >> real :: result >> >> call mpi_init(ierr) >> call mpi_comm_size(mpi_comm_world, size, ierr) >> call mpi_comm_rank(mpi_comm_world, rank, ierr) >> >> ! create groups >> call mpi_comm_group(mpi_comm_world, group_all, ierr) >> >> if (rank < 4 ) then >> group_members = (/0,1,2,3/) >> else >> group_members = (/4,5,6,7/) >> endif >> >> call mpi_group_incl(group_all, group_size, group_members, subgroup, ierr) >> call mpi_comm_create(mpi_comm_world, subgroup, mpi_comm_grid, ierr) >> call mpi_comm_rank(mpi_comm_grid, local_rank, ierr) ! rank within group >> >> ! create window >> call mpi_type_size(mpi_real, bytesize, ierr) >> window_size = bytesize ! one element in the window >> aa = malloc(1) >> call MPI_ALLOC_MEM(window_size, mpi_info_null, aa, ierr) >> >> duplicate(1) = rank >> >> call mpi_win_create(duplicate, window_size, bytesize, MPI_INFO_NULL, >> mpi_comm_grid, win, ierr) >> >> ! grabbing data from rank above >> if (rank < 3 ) then >> owner = rank + 1 >> elseif (rank < 7) then >> owner = rank + 1 - 4 >> else >> owner = 0 >> endif >> >> print*, 'rank owner', rank, owner >> >> target_disp = 0 >> >> call mpi_win_lock(MPI_LOCK_SHARED, owner, 0, win, ierr) >> call mpi_get(result, 1, mpi_real, owner, target_disp, 1, mpi_real, win, >> ierr) >> call mpi_win_unlock(owner, win, ierr) >> >> print*, 'rank, result ', rank, result >> >> ! free window >> call mpi_win_free(win, ierr) >> call mpi_free_mem(duplicate, ierr) >> >> call mpi_finalize(ierr) >> >> end program test_group_win >> ---- >> >> >> _______________________________________________ >> discuss mailing list discuss(a)mpich.org >> To manage subscription options or unsubscribe: >> https://lists.mpich.org/mailman/listinfo/discuss > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss >

1 0

Re: [mpich-discuss] mpi windows with groups mpich2
by Rajeev Thakur 28 Apr '14

28 Apr '14

The program runs fine on my Mac laptop. Which version of MPICH are you using? Rajeev On Apr 28, 2014, at 10:41 AM, Helen Kershaw <hkershaw(a)ucar.edu> wrote: > Hi, > > I'm trying to use MPI one-sided communication with sub-communicators. > I've put my test code at the end of the email. It runs on 8 processors, creates two communicators, makes a window on the communicators, then tries to do some one-sided communication. > > I can use groups OR windows without a problem, but when I use them > together (in the test code) I get a coredump from the first 4 tasks. I > think the one-sided communication may be using the local rank in the sub-communicator (mpi_comm_grid) as if it is the global rank in mpi_comm_world. > > I can run the test code ok using open-mpi, but not mpich2. I might be > missing something obvious. > > Thanks for your help, > Helen > > ----- > program test_group_win > > use mpi > > implicit none > > ! mpi > integer :: rank, size, ierr, status(mpi_status_size) > ! groups > integer, parameter :: group_size = 4 > integer :: group_members(group_size) > integer :: group_all ! mpi_comm_world group > integer :: subgroup > integer :: mpi_comm_grid > integer :: local_rank ! rank within subgroup > > > ! window > integer :: win > integer ::bytesize !> size in bytes of each > element in the window > integer(KIND=MPI_ADDRESS_KIND) :: window_size > pointer(aa, duplicate) > real :: duplicate(*) > integer(KIND=MPI_ADDRESS_KIND) :: target_disp !> displacement in window > integer :: owner > real :: result > > call mpi_init(ierr) > call mpi_comm_size(mpi_comm_world, size, ierr) > call mpi_comm_rank(mpi_comm_world, rank, ierr) > > ! create groups > call mpi_comm_group(mpi_comm_world, group_all, ierr) > > if (rank < 4 ) then > group_members = (/0,1,2,3/) > else > group_members = (/4,5,6,7/) > endif > > call mpi_group_incl(group_all, group_size, group_members, subgroup, ierr) > call mpi_comm_create(mpi_comm_world, subgroup, mpi_comm_grid, ierr) > call mpi_comm_rank(mpi_comm_grid, local_rank, ierr) ! rank within group > > ! create window > call mpi_type_size(mpi_real, bytesize, ierr) > window_size = bytesize ! one element in the window > aa = malloc(1) > call MPI_ALLOC_MEM(window_size, mpi_info_null, aa, ierr) > > duplicate(1) = rank > > call mpi_win_create(duplicate, window_size, bytesize, MPI_INFO_NULL, > mpi_comm_grid, win, ierr) > > ! grabbing data from rank above > if (rank < 3 ) then > owner = rank + 1 > elseif (rank < 7) then > owner = rank + 1 - 4 > else > owner = 0 > endif > > print*, 'rank owner', rank, owner > > target_disp = 0 > > call mpi_win_lock(MPI_LOCK_SHARED, owner, 0, win, ierr) > call mpi_get(result, 1, mpi_real, owner, target_disp, 1, mpi_real, win, > ierr) > call mpi_win_unlock(owner, win, ierr) > > print*, 'rank, result ', rank, result > > ! free window > call mpi_win_free(win, ierr) > call mpi_free_mem(duplicate, ierr) > > call mpi_finalize(ierr) > > end program test_group_win > ---- > > > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] Memory leak comparing communicators
by VandeVondele Joost 28 Apr '14

28 Apr '14

Hi Antonio, right now, a runnable example is or full CP2K program, freely available at cp2k.org, but not very practical as a testcase. It seems to happen at any point where we call: CALL mpi_comm_compare ( comm1, comm2, iout, ierr ) I'll see if I can reproduce it running something from the test directory. Joost ________________________________________ From: discuss-bounces(a)mpich.org [discuss-bounces(a)mpich.org] on behalf of "Antonio J. Peña" [apenya(a)mcs.anl.gov] Sent: Monday, April 28, 2014 6:26 PM To: discuss(a)mpich.org Subject: Re: [mpich-discuss] Memory leak comparing communicators Hi Joost, Can you share the smallest fragment of code with which we can reproduce this? Antonio On 04/28/2014 11:25 AM, VandeVondele Joost wrote: > Hi, > > using mpich 3.1 I notice that my application leaks memory. Compiling it with '-O1 -g -fno-omit-frame-pointer -fsanitize=leak' using gcc 4.9 it appears the sources is comparing communicators, for example: > > [0] Direct leak of 38400 byte(s) in 600 object(s) allocated from: > [0] #0 0x7f84298d53a8 in __interceptor_malloc ../../../../gcc/libsanitizer/lsan/lsan_interceptors.cc:66 > [0] #1 0x7f842aabaf6b in MPIR_Group_create /data/vjoost/mpich-3.1/src/mpi/group/grouputil.c:77 > [0] #2 0x7f842ab88314 in MPIR_Comm_group_impl /data/vjoost/mpich-3.1/src/mpi/comm/comm_group.c:44 > [0] #3 0x7f842ab840d3 in PMPI_Comm_compare /data/vjoost/mpich-3.1/src/mpi/comm/comm_compare.c:122 > [0] #4 0x7f842aac8e84 in pmpi_comm_compare_ /data/vjoost/mpich-3.1/src/binding/f77/comm_comparef.c:267 > [0] #5 0x1bfd447 in __message_passing_MOD_mp_comm_compare /data/vjoost/clean/cp2k/cp2k/src/common/message_passing.F:1084 > > I'm not seeing other leaks, so is guess something particular is going on here. Any suggestions on a possible fix ? > > Thanks, > > Joost > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss -- Antonio J. Peña Postdoctoral Appointee Mathematics and Computer Science Division Argonne National Laboratory 9700 South Cass Avenue, Bldg. 240, Of. 3148 Argonne, IL 60439-4847 apenya(a)mcs.anl.gov www.mcs.anl.gov/~apenya _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0