Section: New Results
AD-adjoints of MPI-parallel codes
Participants : Laurent Hascoët, Ala Taftaf, Sri Hari Krishna Narayanan [Argonne National Lab. (Illinois, USA)] .
We have a long-standing collaboration with Argonne National Lab on the question of adjoint AD of message-passing parallel codes. We continued joint development of the Adjoinable-MPI library (AMPI) that provides efficient tangent and adjoint AD for MPI-parallel codes, independently of the AD tool used (now AdolC, dco, OpenAD, Tapenade).
During her PhD work, Ala Taftaf is considering the question of checkpointing applied to the AD-adjoint of an MPI-parallel code. Checkpointing is a memory/runtime tradeoff which is essential for adjoint AD of large codes, in particular parallel codes. However, for MPI codes this question has always been addressed by ad-hoc hand manipulations of the differentiated code, and with no formal assurance of correctness. Ala Taftaf is investigating the assumptions implicitly made during past experiments, to clarify and generalize them. On one hand we propose an extension of the adjoint of MPI point-to-point communication primitives, so that the semantics of an adjoint program is preserved for any placement of checkpoints. On the other hand, we propose an alternative extension of these adjoint communications, more efficient but that requires a number of restrictions on the placement of checkpoints. We shall try to provide proof of correctness of these strategies, and in particular demonstrate that they cannot introduce deadlocks. Tradeoffs between the two extensions should be investigated. Ala Taftaf presented her research on “`Adjoint-Checkpointing on MPI-parallel codes” at the EuroAD workshop in Paderborn, Germany, december 1-2. A conference article has been submitted to Eccomas 2016 in Crete.