A simple way to speed up source association would be to parallelize across ccds within a visit. ccds in a visit will not have any associations between them (no overlap), so that provides a safe parallelization point.

This would require reworking Associations::associateCatalogs to make the ccdImage loop be over visits, with an internal parallelized loop over ccds. The measuredStar->fittedStar association step would be thread safe (because of no overlap between ccds), but the adding of unmatched per-ccd objects to fittedStarList would have to be done in a per-visit gather step.

I don't know how much the parallelization discussion so far has included cases like this; it's pure C++ and mostly trivial (no cross-thread talk except in the per-visit gather).