Dual Rail Performance

David_Race — Thu, 29 Mar 2012 15:04:17 GMT

Hello,

I have a dual rail FDR system with SandyBridge nodes. I have 16 processes assigned to each of four nodes and I am testing the all-to-all. I have enabled dual rail with

export I_MPI_OFA_NUM_ADAPTERS=2

export I_MPI_OFA_RAIL_SCHEDULER=ROUND_ROBIN

The ibstat shows

CA 'mlx4_0'

CA type: MT4099

Number of ports: 1

Firmware version: 2.10.2370

Hardware version: 0

Node GUID: 0x001e6703003dd888

System image GUID: 0x001e6703003dd88b

Port 1:

State: Active

Physical state: LinkUp

Rate: 56

Base lid: 78

LMC: 0

SM lid: 3

Capability mask: 0x02514868

Port GUID: 0x001e6703003dd889

Link layer: InfiniBand

CA 'mlx4_1'

CA type: MT4099

Number of ports: 1

Firmware version: 2.10.700

Hardware version: 0

Node GUID: 0x0002c90300333e90

System image GUID: 0x0002c90300333e93

Port 1:

State: Active

Physical state: LinkUp

Rate: 56

Base lid: 779

LMC: 0

SM lid: 618

Capability mask: 0x02514868

Port GUID: 0x0002c90300333e91

Link layer: InfiniBand

Where does Intel MPI pick up the name of the adapters to use for dual rail mode?

Thanks

David Race

Dual Rail Performance

Dmitry_K_Intel2 — Thu, 29 Mar 2012 15:25:16 GMT

Hi David,
Intel MPI calls an IB verbs function to get a list of available IB devices. So, Intel MPI doesn't work with names of the adapters and doesn't try to read any configuration file. It should be done bylibibvers library.

Regards!
Dmitry

Dual Rail Performance

David_Race — Thu, 29 Mar 2012 17:01:27 GMT

I set I_MPI_DEBUG=10, then received

[16] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1

[16] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2

[28] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1

[28] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2

[32] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1

[32] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2

[35] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1

[35] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2

[36] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1

[36] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2

[37] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1

[37] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2

[38] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1

[40] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1

[40] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2

[41] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1

[41] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2

[42] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1

[42] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2

[43] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1

[43] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2

[48] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1

[48] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2

[16] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1[16] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2[28] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1[28] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2[32] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1[32] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2[35] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1[35] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2[36] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1[36] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2[37] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1[37] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2[38] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1[40] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1[40] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2[41] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1[41] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2[42] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1[42] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2[43] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1[43] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.2[48] DAPL startup(): trying to open default DAPL provider from dat registry: ofa-v2-mlx4_0-1[48] I_MPI_dlopen_dat(): trying to load default dat library: libdat2.so.

From this an the Intel MPI Documentation, does this mean I need an entry for ofa-v2-mlx4_1-1 in the /etc/dat.conf file?

I only have an entry for the first IB devices in the /etc/dat.conf file.

Is this correct?

Thanks

David

Dual Rail Performance

karl_lehnberger — Fri, 30 Mar 2012 09:12:45 GMT

Hi David,

afaik, you cannot use dapl for multirail. You'll have to use the shm:ofa fabric instead.

cheers.

Dual Rail Performance

Dmitry_K_Intel2 — Fri, 30 Mar 2012 09:37:35 GMT

By default Intel MPI uses shm:dapl fabric.
To enable OFA fabric you need to either set environment variable I_MPI_FABRICS=shm:ofa or add an option '-genv I_MPI_FABRICS shm:ofa' to your mpirun command.

Multi-rail feature is available with ofa fabric only and only with versions 4.0.x

Regards!
Dmitry

topic Dual Rail Performance in Intel® MPI Library

Dual Rail Performance

Dual Rail Performance

Dual Rail Performance

Dual Rail Performance

Dual Rail Performance