BibTeX
@ARTICLE{
         Herrmann2020HRA,
       author = "Herrmann, Julien and {Pallez (Aupy)}, Guillaume",
       title = "\textsc{H-Revolve}: A Framework for Adjoint Computation on Synchronous
         Hierarchical Platforms",
       year = "2020",
       publisher = "Association for Computing Machinery",
       address = "New York, NY, USA",
       volume = "46",
       number = "2",
       issn = "0098-3500",
       xurl = "https://doi.org/10.1145/3378672",
       doi = "10.1145/3378672",
       abstract = "We study the problem of checkpointing strategies for adjoint computation on
         synchronous hierarchical platforms, specifically computational platforms with several levels of
         storage with different writing and reading costs. When reversing a large adjoint chain, choosing
         which data to checkpoint and where is a critical decision for the overall performance of the
         computation. We introduce H-Revolve, an optimal algorithm for this problem. We make it available in
         a public Python library along with the implementation of several state-of-the-art algorithms for the
         variant of the problem with two levels of storage. We provide a detailed description of how one can
         use this library in an adjoint computation software in the field of automatic differentiation or
         backpropagation. Finally, we evaluate the performance of H-Revolve and other checkpointing
         heuristics though an extensive campaign of simulation.",
       journal = "{ACM} Transactions on Mathematical Software",
       month = "jun",
       articleno = "12",
       numpages = "25",
       pages = "12:1--12:25",
       keywords = "revolve, hierarchical memory, deep learning, automatic differentiation, Adjoint
         computation",
       ad_theotech = "Checkpointing"
}