Schedules for Dynamic Bidirectional Simulations on Parallel Computers

For adjoint calculations, parameter estimation, and similar purposes one may need to reverse the execution of a computer program. The simplest option is to record a complete execution log and then to read it backwards. This requires massive amounts of storage. Instead one may generate the execution log piecewise by restarting the ``forward'' calculation repeatedly from suitably placed checkpoints. This thesis extends the theoretical results of the parallel reversal schedules. First a algorithm was constructed which carries out the ``forward'' calculation and distributes checkpoints in a way, such that the reversal calculation can be started at any time. This approach provides adaptive parallel reversal schedules for simulations where the number of time steps is not known a-priori. The number of checkpoints and processors used is optimal at any time. Further, an algorithm was described which makes is possible to restart the initial computer program during the program reversal. Again, this can be done without any additional computation at any time. Hence, optimal parallel reversal schedules for the bidirectional simulation are provided by this thesis.

Level thesis:degree_level

thesis.doctoral

Grantor dc:publisher

Technische Universität Dresden

Year

2003

Author dc:creator

• Lehmann, Uwe

Contributors dc:contributor

• Griewank, Andreas
• Nagel, Wolfgang E.
• Gockenbach, Mark S.
• Sørevik, Tor

• Automatisches Differenzieren
• Berechnung von Adjungierten
• Bidirektionale Simulation
• Checkpointing
• Parallel
• Programmumkehr
• Adjoint calculation
• Automatic Differentiation
• Bidirectional Simulations
• Program Reversal