Storing versus recomputation on multiple DAGs

Abstract

Recomputation and storing are typically seen as tradeoffs for checkpointing schemes in the context of adjoint computations. At finer granularity during the adjoint sweep, in practice, only the store-all or recompute-all approaches are fully automated. This paper considers a heuristic approach for exploiting finer granularity recomputations to reduce the storage requirements and thereby improve the overall adjoint efficiency without the need for manual intervention. © 2012 Springer-Verlag.