Adjoinable MPI: From the Theory to a Reusable Implementation
In the context of gradient computations via adjoint mode algorithmic differentiation there have been various ad hoc solutions for simple blocking MPI communications. A few years ago the demand arose for a treatment of nonblocking MPI communication patterns. The first part of the talk will describe the conceptual approach to adjoining the MPI communications.
To make the approach useful for practical applications, implementations have to be developed in the context of existing algorithmic differentiation tools. Like MPI itself, the adjoinable MPI should have an interface that is independent of the implementation context. The second part of the talk will discuss this ongoing effort. The talk will conclude with an outlook at future work on one-sided communications.
Adobe Connect Meeting Information
If you have never attended an Adobe Connect meeting before: