Without more information I can only add some random guesses:
dwiintensitynorm uses population_template which as of this commit has double the number of linear registration stages. I have never run it on such a huge population but it should not increase the computation time by two orders of magnitude.
The dirty pages issue could cause such a decrease in performance but I can not think of any recent changes to the code that might have increased this problem. Can you monitor the CPU and machine load with top. If there is a swapping issue then the load average is high but the CPU wa is low.
What is the old version you are referring to? We changed the behaviour of dwiintensitynorm regarding symbolic links in November.
population_template’s temporary directory is created inside dwiintensitynorm’s. So if this used to be in RAM and now is on disk, this might impact performance. Also, if it is now in RAM and you run out of RAM then swapping will cause this decrease in performance.