Asynchronous Two-Level Checkpointing Scheme for Large-Scale Adjoints using Nek5000

Event Sponsor: 
Mathematics and Computing Science - LANS Seminar
Start Date: 
Nov 4 2015 - 3:00pm
Building 240/Room 4301
Argonne National Laboratory
Michel Schanen
Speaker(s) Title: 
Postdoctoral Appointee, MCS
Kamil Khan

Adjoints are an important computational tool for large-scale sensitivity evaluation, uncertainty quantification, and derivative-based optimization. Essential to their performance is an efficient checkpointing scheme recovering the primal values in the adjoint run, this being a trade-off between memory requirement and recomputation. An asynchronous two-level adjoint checkpointing scheme has been implemented for multistep numerical time discretizations targeted at large-scale numerical simulations. The checkpointing scheme combines both bandwidth-limited disk checkpointing and binomial memory checkpointing. Based on assumptions about the target petascale system, 50k+ cores on the IBM Blue Gene/Q system Mira, this checkpointing approach and the performance model has been validated using the spectral element solver Nek5000.

Miscellaneous Information: 

Coffee and Goodies will be served.

Please click below to add this event to your calendar.