Re: [OMPI users] Conflicting directives for mapping policy

Thu, 4 Feb 2016 02:19:24 -0500 (EST) 

Here is what i get trying something similar :

$ mpirun -np 1 --npernode 1 --map-by core hostname
--------------------------------------------------------------------------
Conflicting directives for mapping policy are causing the policy
to be redefined:

  New policy:   bynode
  Prior policy:  BYCORE

Please check that only one policy is defined.
--------------------------------------------------------------------------
i assume you invoke mpirun from a batch manager, if so, did you also invoke ompi_info from the same job ? 
if not, then i have no idea where RANK_FILE is coming from.
currently, --npernode is telling ompi to use the "ppr" mapper, and it causes a conflict with the "map by core" policy. 

Ralph,
can you please comment on that ?

btw,

i also tried to map by node, and the result is quite surprising :
$ mpirun -np 1 --npernode 1 --map-by node hostname
--------------------------------------------------------------------------
Conflicting directives for mapping policy are causing the policy
to be redefined:

  New policy:   bynode
  Prior policy:  BYNODE

Please check that only one policy is defined.
--------------------------------------------------------------------------

Cheers,

Gilles

On 2/4/2016 1:59 PM, Jingchao Zhang wrote:


Hi Gilles,


Here is what I have for v1.8.8

$ompi_info --all | grep rmaps_rank_file
514: MCA orte: parameter "orte_rankfile" (current value: "", data source: default, level: 5 tuner/detail, type: string, synonym of: rmaps_rank_file_path) 774: MCA rmaps: parameter "rmaps_rank_file_priority" (current value: "0", data source: default, level: 9 dev/all, type: int) 776: MCA rmaps: parameter "rmaps_rank_file_path" (current value: "", data source: default, level: 5 tuner/detail, type: string, synonyms: orte_rankfile) 778: MCA rmaps: parameter "rmaps_rank_file_physical" (current value: "false", data source: default, level: 5 tuner/detail, type: bool) 

and v1.10.1

$ompi_info --all | grep rmaps_rank_file
514: MCA orte: parameter "orte_rankfile" (current value: "", data source: default, level: 5 tuner/detail, type: string, synonym of: rmaps_rank_file_path) 770: MCA rmaps: parameter "rmaps_rank_file_priority" (current value: "0", data source: default, level: 9 dev/all, type: int) 772: MCA rmaps: parameter "rmaps_rank_file_path" (current value: "", data source: default, level: 5 tuner/detail, type: string, synonyms: orte_rankfile) 774: MCA rmaps: parameter "rmaps_rank_file_physical" (current value: "false", data source: default, level: 5 tuner/detail, type: bool)
I am not sure if this is the root of the problem because v1.8.4 seems to have the same results 

$ompi_info --all | grep rmaps_rank_file
508: MCA orte: parameter "orte_rankfile" (current value: "", data source: default, level: 5 tuner/detail, type: string, synonym of: rmaps_rank_file_path) 755: MCA rmaps: parameter "rmaps_rank_file_priority" (current value: "0", data source: default, level: 9 dev/all, type: int) 757: MCA rmaps: parameter "rmaps_rank_file_path" (current value: "", data source: default, level: 5 tuner/detail, type: string, synonyms: orte_rankfile) 759: MCA rmaps: parameter "rmaps_rank_file_physical" (current value: "false", data source: default, level: 5 tuner/detail, type: bool) 

Thanks,


Dr. Jingchao Zhang
Holland Computing Center
University of Nebraska-Lincoln
402-472-6400


------------------------------------------------------------------------
*From:* users <users-boun...@open-mpi.org> on behalf of Gilles Gouaillardet <gil...@rist.or.jp> 
*Sent:* Wednesday, February 3, 2016 5:44 PM
*To:* Open MPI Users
*Subject:* Re: [OMPI users] Conflicting directives for mapping policy
I checked the v1.10 source code, and for some reasons, it seems
rmaps_rank_file_path is defined.

before invoking mpirun, can you please run
ompi_info --all | grep rmaps_rank_file

and confirm that

Cheers,

Gilles

On 2/4/2016 2:07 AM, Jingchao Zhang wrote:


Hi all,
We have an application named "GAMESS" that worked fine with OpenMPI v1.8.4 but having problems with v1.8.8. 


The mpi command in the application is:
orterun --oversubscribe -mca hwloc_base_binding_policy core:overload-allowed -np $NPROCS --npernode $PPN2 $GMSPATH/gamess.$VERNO.x < /dev/null
The above command works for v1.8.4 but for v1.8.8, it give the following error: 

--------------------------------------------------------------------------
Conflicting directives for mapping policy are causing the policy
to be redefined:

  New policy:   RANK_FILE
  Prior policy:  BYCORE:OVERSUBSCRIBE

Please check that only one policy is defined.
--------------------------------------------------------------------------
We used to have binding issues with v1.8.8 which led to poor application performances. To resolve it, we added the following parameters to the "openmpi-mca-params.conf" file: 

orte_hetero_nodes=1
hwloc_base_binding_policy=core
rmaps_base_mapping_policy=core
The above changes in v1.8.8 work great for other stuff but breaks GAMESS. Does anyone know how to resolve the conflict? Any suggestion is appreciated. 


Thanks,


Dr. Jingchao Zhang
Holland Computing Center
University of Nebraska-Lincoln
402-472-6400


_______________________________________________
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/2016/02/28444.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/2016/02/28446.php

Reply via email to