Closed fwinter closed 12 years ago
Hi,
running on 2 hosts with 1 mpi node per host does not work for me. Whereas running on 1 host with 2 mpi nodes works fine. I tracked it down to
face_qmp.cpp: void FaceBuffer::exchangeCpuLink(void* ghost_link, void* link_sendbuf)
where the statement
for (int i=0; i<4; i++) { QMP_wait(mh_send_fwd[i]); QMP_wait(mh_from_back[i]); }
never completes. It just waits forever.
I use a QMP enabled build. Machine type: 4xGTX480 (but hostfile has just 1 slot per host). Quda commit is 5c24b4556fc7f374bfd
As I said, running the same QMP build on 1 host with 2 slots per host just runs fine.
Any ideas?
Frank, Can you try and narrow down to which commit it breaks? The commit you tried has bunch of MPI related changes but it should not affect qmp code path.
-Guochun
Okay, current master fixes this.
Hi,
running on 2 hosts with 1 mpi node per host does not work for me. Whereas running on 1 host with 2 mpi nodes works fine. I tracked it down to
face_qmp.cpp: void FaceBuffer::exchangeCpuLink(void* ghost_link, void* link_sendbuf)
where the statement
for (int i=0; i<4; i++) { QMP_wait(mh_send_fwd[i]); QMP_wait(mh_from_back[i]); }
never completes. It just waits forever.
I use a QMP enabled build. Machine type: 4xGTX480 (but hostfile has just 1 slot per host). Quda commit is 5c24b4556fc7f374bfd
As I said, running the same QMP build on 1 host with 2 slots per host just runs fine.
Any ideas?