Open MPI logo

Open MPI User's Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Open MPI User's mailing list

Subject: [OMPI users] Selective file driven decision functions for Open MPI coll tuned
From: Norio Tamura (tam_at_[hidden])
Date: 2009-03-16 21:55:29

Hello everyone,

Could anybody let me know where I can find a manual or an web page that describes
the specification for specifying the selective file driven decision functions
in Open MPI's mpirun/mpiexec command as explained below ?

> Jelena Pjesivac-Grbovic wrote:
>This is an excellent solution if we have only one type of allreduce (large
>one). However, if your application has multiple types of allreduce (both
>small and large), this may not be a satisfactory solution. In this case,
>Open MPI allows you to specify a file which contains rules which function
>to use when. This option has slight learning curve and I am currently
>evaluating its effectiveness, but if you are interested - I will send you
>more documentation about it.

>Selective file driven decision functions Another alternative to forcing the
>complete collective operations is to force only parts of the decision space in a
>semi-fixed manner. An example of such a usage scenario would be in the case of
>a user having tuned an MPI collective for a range of input parameters (message
>size, communicator size) either manually or via some automated tool [6]. The
>user could then tell the MPI collective component to use these values within a
>set range by supplying a file that contains as many data points as the user knows.
>To decrease both storage and evaluation time the file contents are stored using
>a run-length technique that effectively only stores the switching points for each
>algorithm. An example version for an MPI Alltoall operation is shown below:
>1 # num of collectives
>3 # ID = 3 Alltoall collective (ID in coll_tuned.h)
>2 # number of com sizes
>1 # comm size 1
>1 # number of msg sizes 1
>0 1 0 0 # for message size 0, linear 1, topo 0, 0 segmentation
>8 # comm size 8
>4 # number of msg sizes
>0 1 0 0 # for message size 0, linear 1, topo 0, 0 segmentation
>32768 2 0 0 # 32k, pairwise 2, no topo or segmentation
>262144 1 0 0 # 256k, use linear 1, no topo or segmentation
>524288 2 0 0 # message size 512k+, pairwise 2, topo 0, 0 segmentation
># end of first collective

Best regards,