Section: New Software and Platforms

disk-revolve

Keywords: Automatic differentiation - Gradients - Machine learning

Functional Description: This software provides several algorithms (Disk-Revolve, 1D-Revolve, Periodic-Disk-Revolve,...) computing the optimal checkpointing strategy when executing a adjoin chain with limited memory. The considered architecture has a level of limited memory that is free to access (writing and reading costs are negligible) and a level of unlimited memory with non-negligible access costs. The algorithms describe which data should be saved in the memory to minimize the number of re-computation during the execution.

• Authors: Guillaume Aupy and Julien Herrmann

• Contact: Julien Herrmann

• Publications: H-Revolve: A Framework for Adjoint Computation on Synchrone Hierarchical Platforms - Periodicity in optimal hierarchical checkpointing schemes for adjoint computations - Optimal Multistage Algorithm for Adjoint Computation

• URL: https://gitlab.inria.fr/adjoint-computation/disk-revolve-public