afni.nimh.nih.gov

As an update to this, I tested reducing OMP_NUM_THREADS down from 7, to 3,2 and 1. The original example above used 7 threads, and had 4 apparent load errors. With 3 and 2 threads I still had 2 and 1 errors respectively. Turning off multi-threading (the 1 case) led to no errors whatsoever (but took forever, of course).

So if the error matters, it appears to be dependent on the number of parallel threads used.