You are here

"abinitio" with mpi

2 posts / 0 new
Last post
"abinitio" with mpi

I have looked at both documentation and this forum, and found no clear answers. I would appreciate it if someone could help.

I have compiled the mpi version to run my "ab initio" structure predictions on multiple cores.  I am not sure if I have to change my flags (such as -jran and -constant_seed etc) and my program stalls at protocols.jobdist.JobDistributors: (0) Master Node -- Waiting for job request; tag_ = 1

I use mpirun -np 4 path/to/AbinitioRelax.mpi.linuxgccrelease @flag

Thanks, any help appreciated.


Post Situation: 
Wed, 2016-10-12 12:17

Do things run appropriately (with small -nstruct) if you run it with non-MPI? (Either with AbinitioRelax.default.linuxgccrelease, or even with AbinitioRelax.mpi.linuxgccrelease  when run directly without the mpirun command.) Generally speaking, it's good to debug jobs in serial mode (with small number of output structures) before scaling things up with mpi running.

I'd also double check the documentation for your MPI system - make sure that mpirun is the launcher you need to be using with the MPI libraries you compiled Rosetta with, and that -np is the appropriate flag to use. (Unfortunately, each MPI system and every cluster seems to have their own quirks on how to properly launch mpi runs.)

At any rate, -constant_seed and -jran definitely *won't* affect the ability to run under MPI.

One thing to try when debugging running with MPI is to add somethign like "-mpi_tracer_to_file mpi_log" to get each processor's tracer output in different files. (To see what errors are coming from which nodes.)

Fri, 2016-10-14 15:46