Re: [mpich-discuss] Issue with OrangeFS 2.9.7 direct interface and MPICH 3.3.1 using CH4 device
Hi Rob, The liborangefsposix library works when I use standard file path like /mnt/pfs/data instead of pvfs2://mnt/pfs/data. I want to use that because it gives better performance in metadata operation which I need for my project. I'm not using MPI I/O to access the files on OrangeFS. It is just a standard C interface call (fopen and fclose). CH4 and liborangefsposix is the only combination that produces this error. Here is the result matrix I have tried. MPICH device Linking OrangeFS Result CH3 non direct interface success CH3 direct interface success CH4 non direct interface success CH4 direct interface no output Thanks Kun On Tue, Oct 8, 2019 at 10:10 AM Latham, Robert J. <[email protected]> wrote:
On Sun, 2019-10-06 at 12:20 -0500, Kun Feng via discuss wrote:
Hi Min,
If that is the case, please ignore this email. Nothing is wrong without OrangeFS direct interface. I will try "ch4:ucx". Thank you for the info.
Does the 'pvfs2' driver still work? the liborangefsposix library might be intercepting system calls MPICH expects to use natively.
The liborangefsposix library is intended more for non-mpi applications -- Hadoop workflows, for example. MPICH's pvfs2 driver (the old name for OrangeFS) speaks directly to the orangefs servers. It also uses a few optimizations not available if MPICH treats the OrangeFS like a traditional UNIX-like file system.
==rob
On Sun, Oct 6, 2019 at 10:25 AM Si, Min via discuss < [email protected]> wrote:
Hi Kun,
Can you please try to reproduce the issue in a simple MPI program which does not use OrangeFS ? It is hard for the MPICH community to help when mixing MPI and OrangeFS together, because we are not OrangeFS experts.
Besides, for InfiniBand networks, you might want to use `ch4:ucx` instead of `ch4:ofi`. But I do not think it causes the failure in your use case.
Best regards, Min
On 2019/10/04 12:21, Kun Feng via discuss wrote:
To Whom It May Concern,
Recently, I switched to CH4 device in MPICH 3.3.1 for better network performance over the RoCE network we are using. I realized that my code fails to run when I use direct interface of OrangeFS 2.9.7. It exits without any error. But even simple helloworld cannot print anything. It happens only when I enable direct interface of OrangeFS by linking -lorangefsposix. Could you please help me on this issue? Here are some information that might be useful: Output of ibv_devinfo of 40Gbps Mellanox ConnectX-4 Lx adapter: hca_id: mlx5_0 transport: InfiniBand (0) fw_ver: 14.20.1030 node_guid: 248a:0703:0015:a800 sys_image_guid: 248a:0703:0015:a800 vendor_id: 0x02c9 vendor_part_id: 4117 hw_ver: 0x0 board_id: LNV2430110027 phys_port_cnt: 1 port: 1 state: PORT_ACTIVE (4) max_mtu: 4096 (5) active_mtu: 1024 (3) sm_lid: 0 port_lid: 0 port_lmc: 0x00 link_layer: Ethernet
hca_id: i40iw0 transport: iWARP (1) fw_ver: 0.2 node_guid: 7cd3:0aef:3da0:0000 sys_image_guid: 7cd3:0aef:3da0:0000 vendor_id: 0x8086 vendor_part_id: 14289 hw_ver: 0x0 board_id: I40IW Board ID phys_port_cnt: 1 port: 1 state: PORT_ACTIVE (4) max_mtu: 4096 (5) active_mtu: 1024 (3) sm_lid: 0 port_lid: 1 port_lmc: 0x00 link_layer: Ethernet MPICH 3.3.1 configuration command: ./configure --with- device=ch4:ofi --with-pvfs2=/home/kfeng/install --enable-shared --enable-romio --with-file-system=ufs+pvfs2+zoidfs --enable- fortran=no --with-libfabric=/home/kfeng/install OrangeFS 2.9.7 configuration command: ./configure -- prefix=/home/kfeng/install --enable-shared --enable-jni --with- jdk=/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.191.b12-0.el7_5.x86_64 --with-kernel=/usr/src/kernels/3.10.0-862.el7.x86_64 Make command: mpicc -o ~/hello ~/hello.c -L/home/kfeng/install/lib -lorangefsposix The verbose outputs of mpiexec are attached.
Thanks Kun
_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss
_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss
_______________________________________________ discuss mailing list [email protected] To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss
participants (1)
-
Kun Feng