Discrete adjoints for MPI-parallelized C++ models with an application to the NASA/JPL Ice Sheet System Model

Dr. Jean Utke, Argonne National Laboratory, U.S.A.

6 Feb 2014, 17:00–18:30; Location: S4|10-1

Computing discrete adjoints by algorithmic differentiation (AD) enables gradient-based optimization for high-dimensional problems. First, I introduce AD principles and implementation options (i.e. operator overloading) relevant to models written in C++ and parallelized with MPI. Then I present our ongoing work with the Ice Sheet System Model (ISSM), developed at NASA/JPL and UC Irvine and used by cryosphere scientists to project the future evolution of polar ice caps such as Greenland or Antarctica. Most of the model-specific AD effort relates to facilitating a type change (for the operator overloading) that is transparent to the developers, particularly to outside contributors unaware of the adjoint capabilities. Two important aspects I discuss are the binding to (external) solvers and the use of the AdjoinableMPI wrapper library to cover adjoining the MPI communication. The presentation will conclude with some performance results showing sources of overhead for the adjoint and options to mitigate them.

Category: CE Seminar


Technische Universität Darmstadt

Graduate School CE
Dolivostraße 15
D-64293 Darmstadt

Phone+49 6151/16-24401
Fax -24404

to assistants' office

Open BSc/MSc Theses

Show a list of open BSc/MSc topics at GSC CE.

 Print |  Impressum |  Sitemap |  Search |  Contact |  Privacy Policy
zum Seitenanfangzum Seitenanfang