This looks like a relatively old paper (2002) and so I think it is measuring performance on strictly single processor machines. On the other hand it is trie based which suggests that it could also be highly parallelizable on modern multi-processor systems. (I am thinking clojure data structures here and maybe the submitter was too... http://fogus.me).
Then again maybe datasets of the size tested (or much bigger) are now just broken down in a map-reduce fashion over multiple machines, and then who cares about the efficiencies of cache misses etc?
Can anyone comment on if burstsort or trie based sorting algorithms in general have been adopted? (All my experience has been as a consumer of library provided "sort" methods rather than ever looking at current state of the art algorithms).