There is a whole page on valgrind web page about this topic. Please read http://valgrind.org/docs/manual/manual-core.html#manual-core.suppress for more information.

  george.

On Jun 8, 2009, at 15:24 , Ralph Castain wrote:

We deliberately choose to not initialize our msg buffers as this takes considerable time. Instead, we fill in only the portion required by a given message, and then send only that much of the buffer. Thus, the uninitialized portion is ignored.

I don't know of a way to tell valgrind to ignore it, I'm afraid - perhaps a valgrind guru can be of help. :-/

Ralph


On Mon, Jun 8, 2009 at 1:09 PM, tom fogal <tfo...@alumni.unh.edu> wrote:
Hi all,

I've configured a source build of OpenMPI 1.3.2 with valgrind enabled
[1], and I'm seeing a lot of errors with writev() when I run this under
valgrind.  For example, with the following `hello, world' program:

 #include <stdio.h>
 #include <mpi.h>

 int main(int argc, char *argv[]) {
   MPI_Init(&argc, &argv);

   puts("Hello, world!");
   MPI_Finalize();
   return 0;
 }

I see errors like the following:

==12342== Syscall param writev(vector[...]) points to uninitialised byte(s)
 ==12342==    at 0x61DF733: writev (in /lib/libc-2.7.so)
==12342== by 0x7889AB9: mca_oob_tcp_msg_send_handler (oob_tcp_msg.c:265)
 ==12342==    by 0x788B1A0: mca_oob_tcp_peer_send (oob_tcp_peer.c:197)
 ==12342==    by 0x788FF2A: mca_oob_tcp_send_nb (oob_tcp_send.c:167)
 ==12342==    by 0x767C7EC: orte_rml_oob_send (rml_oob_send.c:137)
==12342== by 0x767D19A: orte_rml_oob_send_buffer (rml_oob_send.c: 269)
 ==12342==    by 0x7C9F3DF: allgather (grpcomm_bad_module.c:369)
 ==12342==    by 0x7C9FD9E: modex (grpcomm_bad_module.c:497)
 ==12342==    by 0x4E6DCAF: ompi_mpi_init (ompi_mpi_init.c:626)

The full vg log is appended [2].  Of course, I could just suppress
this error, but I get this for a lot (every?) MPI call which does
communication, it seems (broadcasts, sends, recv's, allgathers, etc.).
I'm worried a suppression would suppress too much / suppress an error
I've caused.

Have others seen this?  Can I suppress perhaps from the
orte_rml_oob_send_buffer down (safely)?

-tom

[1] configured via: gnu_pkg \
   --enable-debug \
   --enable-memchecker \
   --disable-mpi-f77 \
   --enable-pretty-print-stacktrace \
   --enable-cxx-exceptions \
   --enable-mpi-threads \
   --with-valgrind=${PREFIX} \
   --without-gm \
   --without-mx \
   --without-openib \
   --without-psm \
   --with-pic \
   --with-gnu-ld
 where gnu_pkg is basically a function which calls configure with
 --prefix=${PREFIX}.

[2]
==12342== Memcheck, a memory error detector.
==12342== Copyright (C) 2002-2008, and GNU GPL'd, by Julian Seward et al. ==12342== Using LibVEX rev 1884, a library for dynamic binary translation.
==12342== Copyright (C) 2004-2008, and GNU GPL'd, by OpenWorks LLP.
==12342== Using valgrind-3.4.1, a dynamic binary instrumentation framework. ==12342== Copyright (C) 2000-2008, and GNU GPL'd, by Julian Seward et al.
==12342== For more details, rerun with: -v
==12342==
==12342== My PID = 12342, parent PID = 12341.  Prog and args are:
==12342==    ./a.out
==12342==
==12342== Warning: client syscall munmap tried to modify addresses 0xffffffffffffffff-0xffe ==12342== Syscall param writev(vector[...]) points to uninitialised byte(s)
==12342==    at 0x61DF733: writev (in /lib/libc-2.7.so)
==12342== by 0x7889AB9: mca_oob_tcp_msg_send_handler (oob_tcp_msg.c:265)
==12342==    by 0x788B1A0: mca_oob_tcp_peer_send (oob_tcp_peer.c:197)
==12342==    by 0x788FF2A: mca_oob_tcp_send_nb (oob_tcp_send.c:167)
==12342==    by 0x767C7EC: orte_rml_oob_send (rml_oob_send.c:137)
==12342== by 0x767D19A: orte_rml_oob_send_buffer (rml_oob_send.c: 269)
==12342==    by 0x7C9F3DF: allgather (grpcomm_bad_module.c:369)
==12342==    by 0x7C9FD9E: modex (grpcomm_bad_module.c:497)
==12342==    by 0x4E6DCAF: ompi_mpi_init (ompi_mpi_init.c:626)
==12342==    by 0x4EAAC88: PMPI_Init (pinit.c:80)
==12342==    by 0x400857: main (hello.c:5)
==12342== Address 0x677697b is 107 bytes inside a block of size 256 alloc'd
==12342==    at 0x4C22A51: realloc (vg_replace_malloc.c:429)
==12342== by 0x53DCBE0: opal_dss_buffer_extend (dss_internal_functions.c:63) ==12342== by 0x53DE4BA: opal_dss_copy_payload (dss_load_unload.c: 164)
==12342==    by 0x7C9F314: allgather (grpcomm_bad_module.c:363)
==12342==    by 0x7C9FD9E: modex (grpcomm_bad_module.c:497)
==12342==    by 0x4E6DCAF: ompi_mpi_init (ompi_mpi_init.c:626)
==12342==    by 0x4EAAC88: PMPI_Init (pinit.c:80)
==12342==    by 0x400857: main (hello.c:5)
==12342==  Uninitialised value was created by a stack allocation
==12342==    at 0x53FFA60: opal_ifinit (if.c:147)
{
  <insert a suppression name here>
  Memcheck:Param
  writev(vector[...])
  fun:writev
  fun:mca_oob_tcp_msg_send_handler
  fun:mca_oob_tcp_peer_send
  fun:mca_oob_tcp_send_nb
  fun:orte_rml_oob_send
  fun:orte_rml_oob_send_buffer
  fun:allgather
  fun:modex
  fun:ompi_mpi_init
  fun:PMPI_Init
  fun:main
}
==12342==
==12342== ERROR SUMMARY: 1 errors from 1 contexts (suppressed: 307 from 3)
==12342== malloc/free: in use at exit: 204,012 bytes in 2,022 blocks.
==12342== malloc/free: 10,382 allocs, 8,360 frees, 14,603,162 bytes allocated.
==12342== For a detailed leak analysis,  rerun with: --leak-check=yes
==12342== For counts of detected errors, rerun with: -v
_______________________________________________
users mailing list
us...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/users

_______________________________________________
users mailing list
us...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/users

Reply via email to