False-positive breakdown detection in the iterative solver

[GoogleCodeExporter]

The breakdown detection is not scale-invariant. So, e.g.
./adda -m 5.19 2.79 -shape ellipsoid 0.95 0.9 -eq_rad 4 -lambda 8.565
works fine, while
./adda -m 5.19 2.79 -shape ellipsoid 0.95 0.9 -eq_rad 4000 -lambda 8565
fails with 
ERROR: (../iterative.c:747) QMR_CS fails: |vT.v|/(b.b) is out of bounds (1e-11).

Either the breakdown detection should be updated (by some constant factor, 
based on the right-hand side) or the whole DDA formulation (linear system) 
should be updated, so that the right-hand side is always of order 1. The latter 
seem to be connected with issue 36.

Original issue reported on code.google.com by yurkin on 6 Mar 2013 at 2:14

[GoogleCodeExporter]

This issue was reported by Michele Galletti at 
https://groups.google.com/d/topic/adda-discuss/4qGnS15G4ao/discussion

Original comment by yurkin on 6 Mar 2013 at 3:13

[GoogleCodeExporter]

This issue was closed by revision r1213.

Original comment by yurkin on 30 Mar 2013 at 9:55

• Changed state: Fixed

[GoogleCodeExporter]

Interesting that this problem was present only for QMR_CS iterative solver. Now 
it is fixed and all iterative solvers are scale-invariant in the range 
(approximately) from 1e-100 to 1e100.

Original comment by yurkin on 30 Mar 2013 at 9:58