Let me send you a patch off list that will print out some extra information to see if we can figure out where things are going wrong. We basically depend on the information reported by hwloc so the patch will print out some extra information to see if we are getting good data from hwloc.
Thanks, Rolf >-----Original Message----- >From: users [mailto:users-boun...@open-mpi.org] On Behalf Of Marcin >Krotkiewski >Sent: Friday, August 28, 2015 12:13 PM >To: Open MPI Users >Subject: Re: [OMPI users] Wrong distance calculations in multi-rail setup? > > >Brilliant! Thank you, Rolf. This works: all ranks have reported using the >expected port number, and performance is twice of what I was observing >before :) > >I can certainly live with this workaround, but I will be happy to do some >debugging to find the problem. If you tell me what is needed / where I can >look, I could help to find the issue. > >Thanks a lot! > >Marcin > > >On 08/28/2015 05:28 PM, Rolf vandeVaart wrote: >> I am not sure why the distances are being computed as you are seeing. I do >not have a dual rail card system to reproduce with. However, short term, I >think you could get what you want by running like the following. The first >argument tells the selection logic to ignore locality, so both cards will be >available to all ranks. Then, using the application specific notation you can >pick >the exact port for each rank. >> >> Something like: >> mpirun -gmca btl_openib_ignore_locality -np 1 --mca >> btl_openib_if_include mlx4_0:1 a.out : -np 1 --mca >> btl_openib_if_include mlx4_0:2 a.out : -np 1 --mca >> btl_openib_if_include mlx4_1:1 a.out : --mca btl_openib_if_include >> mlx4_1:2 a.out >> >> Kind of messy, but that is the general idea. >> >> Rolf >>> -----Original Message----- >>> From: users [mailto:users-boun...@open-mpi.org] On Behalf Of >>> marcin.krotkiewski >>> Sent: Friday, August 28, 2015 10:49 AM >>> To: us...@open-mpi.org >>> Subject: [OMPI users] Wrong distance calculations in multi-rail setup? >>> >>> I have a 4-socket machine with two dual-port Infiniband cards >>> (devices >>> mlx4_0 and mlx4_1). The cards are conneted to PCI slots of different >>> CPUs (I hope..), both ports are active on both cards, everything >>> connected to the same physical network. >>> >>> I use openmpi-1.10.0 and run the IBM-MPI1 benchmark with 4 MPI ranks >>> bound to the 4 sockets, hoping to use both IB cards (and both ports): >>> >>> mpirun --map-by socket --bind-to core -np 4 --mca btl >>> openib,self --mca btl_openib_if_include mlx4_0,mlx4_1 ./IMB-MPI1 >>> SendRecv >>> >>> but OpenMPI refuses to use the mlx4_1 device >>> >>> [node1.local:28265] [rank=0] openib: skipping device mlx4_1; it >>> is too far away >>> [ the same for other ranks ] >>> >>> This is confusing, since I have read that OpenMPI automatically uses >>> a closer HCA, so at least some (>=one) rank should choose mlx4_1. I >>> use binding by socket, here is the reported map: >>> >>> [node1.local:28263] MCW rank 2 bound to socket 2[core 24[hwt 0]]: >>> >[./././././././././././.][./././././././././././.][B/././././././././././.][./././././././././. >>> /./.] >>> [node1.local:28263] MCW rank 3 bound to socket 3[core 36[hwt 0]]: >>> >[./././././././././././.][./././././././././././.][./././././././././././.][B/././././././././. >>> /./.] >>> [node1.local:28263] MCW rank 0 bound to socket 0[core 0[hwt 0]]: >>> >[B/././././././././././.][./././././././././././.][./././././././././././.][./././././././././. >>> /./.] >>> [node1.local:28263] MCW rank 1 bound to socket 1[core 12[hwt 0]]: >>> >[./././././././././././.][B/././././././././././.][./././././././././././.][./././././././././. >>> /./.] >>> >>> To check what's going on I have modified btl_openib_component.c to >>> print the computed distances. >>> >>> opal_output_verbose(1, >ompi_btl_base_framework.framework_output, >>> "[rank=%d] openib: device %d/%d distance >>> %lf", ORTE_PROC_MY_NAME->vpid, >>> (int)i, (int)num_devs, >>> (double)dev_sorted[i].distance); >>> >>> Here is what I get: >>> >>> [node1.local:28265] [rank=0] openib: device 0/2 distance 0.000000 >>> [node1.local:28266] [rank=1] openib: device 0/2 distance 0.000000 >>> [node1.local:28267] [rank=2] openib: device 0/2 distance 0.000000 >>> [node1.local:28268] [rank=3] openib: device 0/2 distance 0.000000 >>> [node1.local:28265] [rank=0] openib: device 1/2 distance 2.100000 >>> [node1.local:28266] [rank=1] openib: device 1/2 distance 1.000000 >>> [node1.local:28267] [rank=2] openib: device 1/2 distance 2.100000 >>> [node1.local:28268] [rank=3] openib: device 1/2 distance >>> 2.100000 >>> >>> So the computed distance for mlx4_0 is 0 on all ranks. I believe this >>> should not be so. The distance should be smaller on 1 rank and larger >>> for 3 others, as is the case for mlx4_1. Looks like a bug? >>> >>> Another question is, In my configuration two ranks will have a 'closer' >>> IB card, but two others will not. Since the correct distance to both >>> devices will likely be equal, which device will they choose, if they >>> do that automatically? I'd rather they didn't both choose mlx4_0.. I >>> guess it would be nice if I could by hand specify the device/port, >>> which should be used by a given MPI rank. Is this (going to be) possible >with OpenMPI? >>> >>> Thanks a lot, >>> >>> Marcin >>> >>> _______________________________________________ >>> users mailing list >>> us...@open-mpi.org >>> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users >>> Link to this post: http://www.open- >>> mpi.org/community/lists/users/2015/08/27503.php >> ---------------------------------------------------------------------- >> ------------- This email message is for the sole use of the intended >> recipient(s) and may contain confidential information. Any >> unauthorized review, use, disclosure or distribution is prohibited. >> If you are not the intended recipient, please contact the sender by >> reply email and destroy all copies of the original message. >> ---------------------------------------------------------------------- >> ------------- _______________________________________________ >> users mailing list >> us...@open-mpi.org >> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users >> Link to this post: >> http://www.open-mpi.org/community/lists/users/2015/08/27504.php > >_______________________________________________ >users mailing list >us...@open-mpi.org >Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users >Link to this post: http://www.open- >mpi.org/community/lists/users/2015/08/27505.php
