Hi everyone,
I have a cluster of 32 nodes with Infiniband, four of them
additionally have a 10G Mellanox Ethernet card for faster I/O. If my
job based on openmpi 1.10.6 ends up on one of these nodes, it will
crash:
No OpenFabrics connection schemes reported that they were able to be
used on a specific port. As such, the openib BTL (OpenFabrics
support) will be disabled for this port.
Local host: pax10-28
Local device: mlx4_0
Local port: 1
CPCs attempted: udcm
--------------------------------------------------------------------------
[pax10-28:08830] *** Process received signal ***
[pax10-28:08830] Signal: Segmentation fault (11)
[pax10-28:08830] Signal code: Address not mapped (1)
[pax10-28:08830] Failing at address: 0x1a0
[pax10-28:08830] [ 0] /usr/lib64/libpthread.so.0(+0xf5e0)[0x2b843325d5e0]
[pax10-28:08830] [ 1]
/opt/ohpc/pub/mpi/openmpi-gnu/1.10.6/lib/openmpi/mca_btl_openib.so(+0x133a0)[0x2b843752e3a0]
[pax10-28:08830] [ 2]
/opt/ohpc/pub/mpi/openmpi-gnu/1.10.6/lib/libopen-pal.so.13(opal_progress+0x2a)[0x2b8433ad1dca]
[pax10-28:08830] [ 3]
/opt/ohpc/pub/mpi/openmpi-gnu/1.10.6/lib/libmpi.so.12(ompi_mpi_init+0x957)[0x2b8432fb4a57]
[pax10-28:08830] [ 4]
/opt/ohpc/pub/mpi/openmpi-gnu/1.10.6/lib/libmpi.so.12(MPI_Init+0x13d)[0x2b8432fd723d]
[pax10-28:08830] [ 5] IMB-MPI1[0x402079]
[pax10-28:08830] [ 6]
/usr/lib64/libc.so.6(__libc_start_main+0xf5)[0x2b843348bc05]
[pax10-28:08830] [ 7] IMB-MPI1[0x401f79]
The same test is working fine if using openmpi 3.0.0, then I get a
similar warning but no crash:
No OpenFabrics connection schemes reported that they were able to be
used on a specific port. As such, the openib BTL (OpenFabrics
support) will be disabled for this port.
Local host: pax10-29
Local device: mlx4_0
Local port: 1
CPCs attempted: rdmacm, udcm
I'm using openmpi packages as included in OpenHPC for Centos 7.
The node providing the error message show two IB devices:
[pax10-28] /root # ibv_devinfo
hca_id: mlx4_1
transport: InfiniBand (0)
fw_ver: 2.33.5100
node_guid: 0cc4:7aff:ff5f:93fc
sys_image_guid: 0cc4:7aff:ff5f:93ff
vendor_id: 0x02c9
vendor_part_id: 4099
hw_ver: 0x0
board_id: SM_2271000001000
phys_port_cnt: 1
port: 1
state: PORT_ACTIVE (4)
max_mtu: 4096 (5)
active_mtu: 4096 (5)
sm_lid: 1
port_lid: 33
port_lmc: 0x00
link_layer: InfiniBand
hca_id: mlx4_0
transport: InfiniBand (0)
fw_ver: 2.36.5000
node_guid: 7cfe:9003:0093:8db0
sys_image_guid: 7cfe:9003:0093:8db0
vendor_id: 0x02c9
vendor_part_id: 4099
hw_ver: 0x0
board_id: MT_1080120023
phys_port_cnt: 2
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
port: 2
state: PORT_DOWN (1)
max_mtu: 4096 (5)
active_mtu: 1024 (3)
sm_lid: 0
port_lid: 0
port_lmc: 0x00
link_layer: Ethernet
The other nodes look like this:
[pax10-27] /root # ibv_devinfo
hca_id: mlx4_0
transport: InfiniBand (0)
fw_ver: 2.33.5100
node_guid: 0cc4:7aff:ff5f:957c
sys_image_guid: 0cc4:7aff:ff5f:957f
vendor_id: 0x02c9
vendor_part_id: 4099
hw_ver: 0x0
board_id: SM_2271000001000
phys_port_cnt: 1
port: 1
state: PORT_ACTIVE (4)
max_mtu: 4096 (5)
active_mtu: 4096 (5)
sm_lid: 1
port_lid: 16
port_lmc: 0x00
link_layer: InfiniBand
So is there a way to tell openmpi to use mlx4_1 on the machines with
10G Ethernet and mlx4_0 on all other nodes?
Regards, Götz
_______________________________________________
users mailing list
users@lists.open-mpi.org
https://lists.open-mpi.org/mailman/listinfo/users
