Section: New Results

Automatic Differentiation and parallel codes

Participants : Valérie Pascual, Laurent Hascoet, Jean Utke [Argonne National Lab. (Illinois, USA)] , Michel Schanen [RWTH Aachen University (Germany)] .

Together with colleagues in Argonne National Lab. and RWTH Aachen, we are studying how AD tools can handle MPI-parallel codes, especially in adjoint mode. Results are progressively incorporated into a library (AMPI, for Adjoinable-MPI) that is designed to provide efficient tangent and adjoint differentiation for MPI-parallel codes, independently of the AD tool used (AdolC, dco, OpenAD, tapenade ...). Primitives from the AMPI library dynamically orchestrate, at run-time, the MPI calls that are needed to compute the derivatives.

This year we studied issues raised by the collective reduction operations of MPI, and by the one-sided communications (i.e. remote memory access) offered by MPI-II.

The participants met on two occasions, two weeks in march in Sophia-Antipolis, and two weeks in october in Argonne.

This work was presented in particular at the meeting of the Inria-Illinois joint lab in june in Lyon. An article is in preparation.