[mvapich-discuss] Error in initializing MVAPICH2 ptmalloc library. Continuing without InfiniBand registration cache support.

Alessandro Fanfarillo fanfarillo.gcc at gmail.com
Tue Jul 5 10:32:20 EDT 2016


Thanks Jonathan, the patch worked perfectly!

Cheers,
Alessandro

2016-07-05 8:12 GMT-06:00 Jonathan Perkins <perkinjo at cse.ohio-state.edu>:
> Hello Alessandro.  It looks like you're facing an issue that a few users
> have hit when using newer versions of gcc.  The following patch (also
> attached) can be applied to resolve the issue for you.  Please apply it
> using patch -p1 <
> 0001-disable-optimization-that-removes-use-of-calloc.patch.  It will be
> available in our next release of MVAPICH2.
>
> commit a1a14d7f35a94d00079d285a7d220ae472afc1d3
> Author: Schulz, Karl W <karl.w.schulz at intel.com>
> Date:   Mon Apr 25 07:31:43 2016 -0400
>
>     disable optimization that removes use of calloc
>
> diff --git a/src/mpid/ch3/channels/common/src/memory/mem_hooks.c
> b/src/mpid/ch3/channels/common/src/memory/mem_hooks.c
> index 26142d2..a52fe4d 100644
> --- a/src/mpid/ch3/channels/common/src/memory/mem_hooks.c
> +++ b/src/mpid/ch3/channels/common/src/memory/mem_hooks.c
> @@ -149,6 +149,12 @@ void mvapich2_mem_unhook(void *ptr, size_t size)
>      }
>  }
>
> +/* disable compiler optimization for minit() to avoid optimizing out memset
> */
> +#ifdef __GNUC__
> +#pragma GCC push_options
> +#pragma GCC optimize ("O0")
> +#endif
> +
>  int mvapich2_minit()
>  {
>      void *ptr_malloc = NULL;
> @@ -214,6 +220,11 @@ int mvapich2_minit()
>      return 0;
>  }
>
> +/* return to default optimization mode */
> +#ifdef __GNUC__
> +#pragma GCC pop_options
> +#endif
> +
>  void mvapich2_mfin()
>  {
>      if (mem_hook_init) {
>
>
>
> On Mon, Jul 4, 2016 at 11:41 PM Alessandro Fanfarillo
> <fanfarillo.gcc at gmail.com> wrote:
>>
>> Dear all,
>> I've installed mvapich2-2.2rc1 on Yellowstone, the NCAR supercomputer,
>> using the following string for the configure:
>>
>> ../configure --with-device=ch3:mrail --with-rdma=gen2
>> --enable-threads=funneled --prefix=/path/to/install
>>
>> Every time I run an MPI program on more than one compute node I get
>> the warning message reported in the subject.
>> I've also tried to set the LD_PRELOAD env var to the libmpi.so as
>> suggested in
>> http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/2015-April/005535.html
>> without any positive effect.
>>
>> Do you have any suggestion on how to fix this problem?
>>
>> Cheers,
>> Alessandro
>> _______________________________________________
>> mvapich-discuss mailing list
>> mvapich-discuss at cse.ohio-state.edu
>> http://mailman.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss


More information about the mvapich-discuss mailing list