On 2014/05/08 2:15, Ralph Castain wrote: > I wonder if that might also explain the issue reported by Gilles regarding > the scif BTL? In his example, the problem only occurred if the message was > split across scif and vader. If so, then it might be that splitting messages > in general is broken. > i am afraid there is a misunderstanding : the problem always occur with scif,vader,self (regardless the ompi v1.8 version) the problem occurs with scif,self only if r31496 is applied to ompi v1.8
In my previous email http://www.open-mpi.org/community/lists/devel/2014/05/14699.php i reported the following interesting fact : with ompi v1.8 (latest r31678), the following command produces incorrect results : mpirun -host localhost -np 2 --mca btl scif,self ./test_scif but with ompi v1.8 r31309, the very same command produces correct results Elena pointed that r31496 is a suspect. so i took the latest v1.8 (r31678) and reverted r31496 and ... mpirun -host localhost -np 2 --mca btl scif,self ./test_scif works again ! note that the "default" mpirun -host localhost -np 2 --mca btl scif,vader,self ./test_scif still produces incorrect results in order to reproduce the issue, a MIC is *not* needed, you only need to install the software stack, load the mic kernel module and make sure you can read/write /dev/mic/* bottom line, there are two issues here : 1) r31496 broke something : mpirun -np 2 -host localhost --mca btl scif,self ./test_scif 2) something else never worked : mpirun -np 2 -host localhost --mca btl scif,vader,self ./test_scif Gilles
