Opened 9 years ago

Closed 2 years ago

#93 closed defect - convergence (fixed)

iterate to convergence, 32 vs 64 bit

Reported by: Gary J. Ferland Owned by: nobody
Priority: major Milestone: C13 branch
Component: radiative transfer Version:
Keywords: Cc:

Description (last modified by Gary J. Ferland)

the default value for the relative error in the iterate to convergence command was changed from 0.2 to 0.1 in r3102 (conv.autocv in zero.cpp). This was to better predict the outward flux in lines, as per the eigenvector 1 paper. Further, conv_itercheck.cpp checks that the outward flux in Hb changes by less than this value.

with this change several blr sims take quite different numbers of iterations to converge on 32 vs 64 bit systems. This change in the default relative error is being reverted in r3154.

the worst offender was blr_n11_p20 - 5 iterations in 64 bit, 9 in 32 bit

Change History (2)

comment:1 Changed 9 years ago by Gary J. Ferland

Description: modified (diff)

comment:2 Changed 2 years ago by peter

Resolution: fixed
Status: newclosed

The convergence has been greatly improved in r8534. The number of iterations in blr_n11_p20 is now 4 on all platforms (at r11136).

Note: See TracTickets for help on using tickets.