Open MPI Development Mailing List Archives

Subject: Re: [OMPI devel] how should processes be mapped to nodes?
From: Eugene Loh (Eugene.Loh_at_[hidden])
Date: 2009-10-08 10:36:01

• Next message: Don Kerr: "Re: [OMPI devel] trac #2034 : single rail openib btl shows better bandwidth than dual rail (12k< x < 128k)"
• Previous message: Eugene Loh: "[OMPI devel] how should processes be mapped to nodes?"
• In reply to: Eugene Loh: "[OMPI devel] how should processes be mapped to nodes?"

Okay, my e-mail below just crossed with another one explaining that
behavior has recently been intentionally changed. I'll go chew on that
for a while.

Eugene Loh wrote:

> If I specify two hosts and no "slots" information, processes should be
> started in equal numbers on the two hosts. But, how should they be
> ranked? "Block" or "cyclic" (round robin)? Did we change this
> behavior recently?
>
>
>
> % ompi_info | grep "Open MPI:"
> Open MPI: 1.3.4a1r22009
> % mpirun -H saem9,saem10 -np 8 -tag-output hostname
> [1,0]<stdout>:saem9
> [1,1]<stdout>:saem9
> [1,2]<stdout>:saem9
> [1,3]<stdout>:saem9
> [1,4]<stdout>:saem10
> [1,5]<stdout>:saem10
> [1,6]<stdout>:saem10
> [1,7]<stdout>:saem10
>
>
>
>
> % ompi_info | grep "Open MPI:"
> Open MPI: 1.3.4a1r22051
> % mpirun -H saem9,saem10 -np 8 -tag-output hostname
> [1,0]<stdout>:saem9
> [1,1]<stdout>:saem10
> [1,2]<stdout>:saem9
> [1,3]<stdout>:saem10
> [1,4]<stdout>:saem9
> [1,5]<stdout>:saem10
> [1,6]<stdout>:saem9
> [1,7]<stdout>:saem10

• Next message: Don Kerr: "Re: [OMPI devel] trac #2034 : single rail openib btl shows better bandwidth than dual rail (12k< x < 128k)"
• Previous message: Eugene Loh: "[OMPI devel] how should processes be mapped to nodes?"
• In reply to: Eugene Loh: "[OMPI devel] how should processes be mapped to nodes?"