Hello, On Saturday 22 August 2009 12:05:00 pm Dr. Hans Ekkehard Plesser wrote: > OpenMPI 1.3.3 comes with a range of algorithms for MPI_Allgather(), but I > have not been able to find out how to check which algorithms is actually > used. You may get this information using the debug output using flag --mca coll_base_verbose 1, however this is turned on only, when build in debugging mode (aka configure --enable-debug), see variable ompi_coll_tuned_stream.
So, e.g. with: mpirun -np 4 --mca coll_base_verbose 1 --mca coll_tuned_use_dynamic_rules 1 Collective tests Allgather (31/1), comm MPI_COMM_WORLD (1/13), type MPI_INT (7/1) [lap79756:05981] coll:tuned:allgather_intra_recursivedoubling rank 0, size 4 [lap79756:05981] ompi_coll_tuned_allgather_intra_dec_dynamic [lap79756:05981] ompi_coll_tuned_allgather_intra_dec_fixed rank 0 com_size 4 msg_length 16000 [lap79756:05982] coll:tuned:allgather_intra_recursivedoubling rank 1, size 4 [lap79756:05982] coll:tuned:allgather_intra_recursivedoubling rank 1, size 4 [lap79756:05982] coll:tuned:module_init called. [lap79756:05982] coll:tuned:module_init MCW & Dynamic [lap79756:05982] coll:tuned:module_init Tuned is in use [lap79756:05982] coll:tuned:module_query using intra_dynamic [lap79756:05982] coll:tuned:module_tuned query called [lap79756:05982] ompi_coll_tuned_allgather_intra_dec_dynamic [lap79756:05982] ompi_coll_tuned_allgather_intra_dec_dynamic [lap79756:05982] ompi_coll_tuned_allgather_intra_dec_fixed rank 1 com_size 4 msg_length 16000 [lap79756:05982] ompi_coll_tuned_allgather_intra_dec_fixed rank 1 com_size 4 msg_length 16000 [lap79756:05984] coll:tuned:allgather_intra_recursivedoubling rank 3, size 4 [lap79756:05984] coll:tuned:allgather_intra_recursivedoubling rank 3, size 4 [lap79756:05984] coll:tuned:module_init called. [lap79756:05984] coll:tuned:module_init MCW & Dynamic [lap79756:05984] coll:tuned:module_init Tuned is in use [lap79756:05984] coll:tuned:module_query using intra_dynamic [lap79756:05984] coll:tuned:module_tuned query called [lap79756:05984] ompi_coll_tuned_allgather_intra_dec_dynamic [lap79756:05984] ompi_coll_tuned_allgather_intra_dec_dynamic [lap79756:05984] ompi_coll_tuned_allgather_intra_dec_fixed rank 3 com_size 4 msg_length 16000 [lap79756:05984] ompi_coll_tuned_allgather_intra_dec_fixed rank 3 com_size 4 msg_length 16000 > From the following message on the list I got an idea about how to select > the algoritm by hand > http://www.open-mpi.org/community/lists/users/2008/10/6847.php, > but I should like to be able to find out at runtime which algorithm Open > MPI has chosen. You want to use this information in Your application? > BTW, is one algorithm chosen the first time MPI_Allgather() is called, or > is a new algorithm chosen for every call based on message properties? If the user does not override the decision, it is done for every MPI_Allgather call. With fixed rules using the tuned collectives, the decision is based on msg. size and whether the number of procs in the communicator is a power of 2. With dynamic tuned collectives on, the application may even override the choice between calls to MPI_Allgathers... George surely will elaborate on the details! With best regards, Rainer PS: More information on the file format for file driven decision functions in http://icl.cs.utk.edu/news_pub/submissions/Flex-collective-euro- pvmmpi-2006.pdf -- ------------------------------------------------------------------------ Rainer Keller, PhD Tel: +1 (865) 241-6293 Oak Ridge National Lab Fax: +1 (865) 241-4811 PO Box 2008 MS 6164 Email: kel...@ornl.gov Oak Ridge, TN 37831-2008 AIM/Skype: rusraink
