I am finding that OpenMPI 1.10.1 will readily enter an infinite loop 
continuously printing

[warn] realloc: Cannot allocate memory

if its virtual address space is severely restricted. The exact point at 
which this occurs will be dependent on the code and compiler, but, as an 
example, a simple "hello world" MPI program compiled with gfortran and run 
with

ulimit -S -v 150000
mpirun ./a.out

triggered this issue on both a dual and a quad core machine. The failure 
with slightly larger amounts of VAS (e.g. 200000) is not very elegant (it 
hangs), but at least it does not then produce an infinite amount of I/O.

The configure options when building OpenMPI were --enable-static 
--disable-shared --disable-ipv6 --disable-heterogeneous 
--enable-mca-no-build=btl-tcp,pls-gridengine,ras-gridengine -without-slurm 
--disable-io-romio --without-loadleveler --without-memory-manager and the 
architecture is x86_64 Linux (OpenSuSE 13.1, own build of OpenMPI). The 
same issue occurs with the Intel compiler suite. The failure does not 
occur when launching non-MPI codes (such as /bin/echo) with mpirun.

Regards,

Michael

Reply via email to