when turning map into hash_map gives you 30.4% increase in performance, you can't call this an 'extensive' tuning effort... The 'extensive' tuning in the paper (structure peeling, ...), accounts for minor improvements (< 10% total).
The initial code was supposed to be written in a straightforward way. That said, I was playing around with the code, and using tcmalloc with the original C++ code gives a better speedup (something close to 40% with the provided example.)
