Currently, even in sparse mode, the table of Sommerfeld integrals is computed
for all possible combinations of dipoles, put in a computational box. So its
size is (approximately) (2*boxZ-1)*(boxX*boxY - MIN(boxX*boxY)^2). This is fine
for FFT mode, but can be very inefficient for SPARSE mode.
1) This procedure incurs (potentially) a lot of unnecessary evaluations
of Sommerfeld integrals. For really sparse aggregates the better way is to
buildup a lookup table, using only actually used pairs of (z,rho), as is done
in DDA-SI code. A hash table can be used for that.
2) Another problem in sparse mode is that currently all values of z are
computed on each processor (in MPI mode). This is related to the current
parallelization mode (issue 160), but it can be improved by computing all
values in chunks and then gathering them on each processor. Better to combine
it with the hash table above.
Original issue reported on code.google.com by yurkin on 25 Sep 2013 at 7:37
Original issue reported on code.google.com by
yurkinon 25 Sep 2013 at 7:37