On 01/20/2016 05:05 PM, Orion Poplawski wrote:
I'm got a strange situation - I'm trying to build hdf5 1.8.16 for Fedora. The Fedora builders do not have network access (/etc/resolv.conf set to "nameserver 127.0.0.1" and no local nameserver) for security purposes. The hdf5 t_mpi parallel test hangs on launch with mpich 3.2, but only on the arm builders.
The processes seem to deadlock in the MPI code, each blocking in poll(), presumably waiting for communication from the other process.
Apparently the Fedora package builds using --with-device=ch3:nemesis on x86 and --with-device=ch3:sock on all other platforms, so the arm builds are using a different channel. Is that still a necessary and/or recommended configuration? -- Orion Poplawski Technical Manager 303-415-9701 x222 NWRA/CoRA Division FAX: 303-415-9702 3380 Mitchell Lane [email protected] Boulder, CO 80301 http://www.cora.nwra.com _______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss