Just a moment:
i didn't compile OpenMPI with bounds checking, but only my application. Problems with OMPI compiled with bound checking remains.
On Jun 19, 2008, at 11:25 AM, Gabriele Fatigati wrote:This is probably not too surprising -- the IB driver gives back memory to the process that didn't come from malloc or other traditional memory allocators, and this can confuse bounds checkers, leading to false positives.
i solved using Gigabit net and eth1 interface!
So, there are problems beetween Infiniband and gcc with bound checking.
In the upcoming v1.3 series, we added a whole pile of stuff to handle these kinds of false positives when running under valgrind. Just curious -- does the gcc bounds checking stuff give you the possibility of saying "this memory is ok"? (valgrind does; we use it for IB-driver allocated or initialized memory)
users mailing list