October 2013 - devel - mailman3.cels.anl.gov

Error check in MPI_Type_get_[true_]extent_x()
by Lisandro Dalcin 15 Oct '13

15 Oct '13

This is a followup on a previous thread in the old mpich-devel ML (cannot find archives of the old list to link here) related to this ticket http://trac.mpich.org/projects/mpich/ticket/1774 In my original report, I commented that the issue affected MPI_Type_size_x() and MPI_Type_get_[true_]extent_x(). However, ticket 1174 was filed only for handling MPI_Type_size_x(). By looking at the source code currently in master, it seems that MPI_Type_get_[true_]extent_x() are coded to fail with uncommitted user-defined datatypes. I think you are still missing the trivial patch below: diff --git a/src/mpi/datatype/type_get_extent_x.c b/src/mpi/datatype/type_get_extent_x.c index 94ba3a4..72f9db1 100644 --- a/src/mpi/datatype/type_get_extent_x.c +++ b/src/mpi/datatype/type_get_extent_x.c @@ -99,7 +99,6 @@ int MPI_Type_get_extent_x(MPI_Datatype datatype, MPI_Count *lb, MPI_Count *exten MPID_Datatype *datatype_ptr = NULL; MPID_Datatype_get_ptr(datatype, datatype_ptr); MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno); - MPID_Datatype_committed_ptr(datatype_ptr, mpi_errno); } /* TODO more checks may be appropriate (counts, in_place, buffer aliasing, etc) */ diff --git a/src/mpi/datatype/type_get_true_extent_x.c b/src/mpi/datatype/type_get_true_extent_x.c index c7b4f8f..4a6d9a9 100644 --- a/src/mpi/datatype/type_get_true_extent_x.c +++ b/src/mpi/datatype/type_get_true_extent_x.c @@ -99,7 +99,6 @@ int MPI_Type_get_true_extent_x(MPI_Datatype datatype, MPI_Count *true_lb, MPI_Co MPID_Datatype *datatype_ptr = NULL; MPID_Datatype_get_ptr(datatype, datatype_ptr); MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno); - MPID_Datatype_committed_ptr(datatype_ptr, mpi_errno); } /* TODO more checks may be appropriate (counts, in_place, buffer aliasing, etc) */ -- Lisandro Dalcin --------------- CIMEC (UNL/CONICET) Predio CONICET-Santa Fe Colectora RN 168 Km 472, Paraje El Pozo 3000 Santa Fe, Argentina Tel: +54-342-4511594 (ext 1016) Tel/Fax: +54-342-4511169

1 0

Re: [mpich-devel] when is MPICH going to be count-safe?
by Rob Latham 14 Oct '13

14 Oct '13

On Mon, Oct 14, 2013 at 02:25:49PM -0500, Rob Latham wrote: > er, maybe I should say "qualified to review, once I pass your tests": > > % ./test_type_contig.x 15 > count = 32768 > ocount = 32768 > % ./test_type_contig.x 25 > count = 33554432 > ocount = 33554432 > % ./test_type_contig.x 30 > count = 1073741824 > ocount = 1073741824 > % ./test_type_contig.x 31 > count = 2147483648 > ocount = -2147483648 > > Heh. silent error vs violent explosion. I'll let you consider which > mode is preferable while I go fix my branch. good news is that I think everything "important" is happening correctly. It's just that when the status object is updated it's picking up an overflowed integer. status.count is MPI_Count big, so that's not the source of the overflow. will look once more at my mountain of clang -Wshorten-64-to-32 related warnings in the morning. ==rob -- Rob Latham Mathematics and Computer Science Division Argonne National Lab, IL USA

1 0

Re: [mpich-devel] when is MPICH going to be count-safe?
by Rob Latham 14 Oct '13

14 Oct '13

On Sun, Oct 13, 2013 at 09:13:31PM -0500, Jeff Hammond wrote: > I'd like to be able to use MPI_Datatypes to send e.g. 4 GiB > messages. Right now, even 2 GiB fails. Right on, man. My "large count large type" branch has been stuck in review purgatory. Think you are you qualified to review mpich-review/large_count_large_typesize_combo ? > Code is here https://github.com/jeffhammond/HPCGloves/tree/master/examples/mpi64. Test machine is in MCS. er, maybe I should say "qualified to review, once I pass your tests": % ./test_type_contig.x 15 count = 32768 ocount = 32768 % ./test_type_contig.x 25 count = 33554432 ocount = 33554432 % ./test_type_contig.x 30 count = 1073741824 ocount = 1073741824 % ./test_type_contig.x 31 count = 2147483648 ocount = -2147483648 Heh. silent error vs violent explosion. I'll let you consider which mode is preferable while I go fix my branch. ==rob -- Rob Latham Mathematics and Computer Science Division Argonne National Lab, IL USA

1 0

Re: [mpich-devel] when is MPICH going to be count-safe?
by Jeff Hammond 13 Oct '13

13 Oct '13

Side note: the OpenMPI trunk gives the correct answers for this test. Jeff ----- Original Message ----- > From: "Jeff Hammond" <jhammond(a)alcf.anl.gov> > To: devel(a)mpich.org > Sent: Sunday, October 13, 2013 9:13:31 PM > Subject: when is MPICH going to be count-safe? > > I'd like to be able to use MPI_Datatypes to send e.g. 4 GiB messages. > Right now, even 2 GiB fails. > > Code is here > https://github.com/jeffhammond/HPCGloves/tree/master/examples/mpi64. > Test machine is in MCS. > > Thanks, > > Jeff > > [jhammond@protos mpi64]$ ./test_type_contig.x 15 > count = 32768 > ocount = 32768 > [jhammond@protos mpi64]$ ./test_type_contig.x 20 > count = 1048576 > ocount = 1048576 > [jhammond@protos mpi64]$ ./test_type_contig.x 25 > count = 33554432 > ocount = 33554432 > [jhammond@protos mpi64]$ ./test_type_contig.x 30 > count = 1073741824 > ocount = 1073741824 > [jhammond@protos mpi64]$ ./test_type_contig.x 31 > count = 2147483648 > Assertion failed in file > ../src/mpid/common/datatype/dataloop/segment_packunpack.c at line > 133: FALSE > memcpy argument memory ranges overlap, dst_=0x254b310 > src_=0x7f6ad81f3010 len_=-2147483648 > > internal ABORT - process 0 > [jhammond@protos mpi64]$ ./test_type_contig.x 32 > count = 4294967296 > ocount = 0 > > -- > Jeff Hammond > Argonne Leadership Computing Facility > University of Chicago Computation Institute > jhammond(a)alcf.anl.gov / (630) 252-5381 > http://www.linkedin.com/in/jeffhammond > https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond > ALCF docs: http://www.alcf.anl.gov/user-guides > -- Jeff Hammond Argonne Leadership Computing Facility University of Chicago Computation Institute jhammond(a)alcf.anl.gov / (630) 252-5381 http://www.linkedin.com/in/jeffhammond https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond ALCF docs: http://www.alcf.anl.gov/user-guides

1 0

when is MPICH going to be count-safe?
by Jeff Hammond 13 Oct '13

13 Oct '13

I'd like to be able to use MPI_Datatypes to send e.g. 4 GiB messages. Right now, even 2 GiB fails. Code is here https://github.com/jeffhammond/HPCGloves/tree/master/examples/mpi64. Test machine is in MCS. Thanks, Jeff [jhammond@protos mpi64]$ ./test_type_contig.x 15 count = 32768 ocount = 32768 [jhammond@protos mpi64]$ ./test_type_contig.x 20 count = 1048576 ocount = 1048576 [jhammond@protos mpi64]$ ./test_type_contig.x 25 count = 33554432 ocount = 33554432 [jhammond@protos mpi64]$ ./test_type_contig.x 30 count = 1073741824 ocount = 1073741824 [jhammond@protos mpi64]$ ./test_type_contig.x 31 count = 2147483648 Assertion failed in file ../src/mpid/common/datatype/dataloop/segment_packunpack.c at line 133: FALSE memcpy argument memory ranges overlap, dst_=0x254b310 src_=0x7f6ad81f3010 len_=-2147483648 internal ABORT - process 0 [jhammond@protos mpi64]$ ./test_type_contig.x 32 count = 4294967296 ocount = 0 -- Jeff Hammond Argonne Leadership Computing Facility University of Chicago Computation Institute jhammond(a)alcf.anl.gov / (630) 252-5381 http://www.linkedin.com/in/jeffhammond https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond ALCF docs: http://www.alcf.anl.gov/user-guides

1 0