[ADOL-C] Speed tapeless vs. tape-based

Werner Jürgen TU Ilmenau juergen.werner at tu-ilmenau.de
Mon Aug 8 08:43:34 EDT 2011


Hi,

I am wondering in which situations a tape-based reverse mode can be expected to be faster than the tapeless forward mode for calculating first derivatives only.

Has anyone tried both on the same problem and can report on the results (run times vs. number of independent variables)?

I'm using the precompiled Windows DLLs (nosparse) from the Paderborn page and the gradient() (or jacobian()) driver. On my actual problem with currently only up to 10 independent variables (expecting to have more later), tape-based is about 5-6 times slower than tapeless. On some toy problem tape-based is slower than tapeless if using up to about 100 independent variables and only faster with more than 100 variables. I was expecting tape-based to be faster than tapeless starting at a significantly lower number of independent variables.

Yours faithfully,
Juergen




More information about the ADOL-C mailing list