Amanda's estimate phase takes here between 3 and 8 hours (with GNU
tar). So I thought about ways of skipping the extra estimate phase.

The estimate does not have to be exactly correct. It really helps
amanda to have most estimates nearly correct, but amanda does not rely
on these numbers.

(I already have directories that are empty at estimate time and the
date arrives there between estimate time and starttime. So I have
bogus estimate values anyway.)

So I think that it might be "good enough" (or even better, see above)
in many case to use the previous run's (same level) size. This assumes
that the size (either of the full dump or of todays differential) does
not change drastically between amanda runs. So the problem is to
decide whether this assumption is correct for the actual data.

This optimization could be selectable per DLE. For DLEs where estimate
is fast enough you don't need this. (But I see no way to decide this
autmatically without really doing the estimate).

Comments? Or what's wrong with my idea?

        Sven