Here is my performance on the 400,000 row dataset on a 2.2GHz AMD Opteron 8354:

cores  runtime  speedup  speedup/cores
 1     25.59s     1        1
 2     12.82s     1.99     .995
 4      6.50s     3.94     .985
 8      3.34s     7.67     .959
12      2.34s    10.96     .913
14      2.02s    12.67     .905
16      1.98s    12.92     .808
c++    23.08s     1.11          (the C++ code with no Cilk++ constructs)

The machine is busy running some other code, so only about 13 or 14 cores are available, hence the falloff in speedup/cores at the top end.

My 1.6GHz laptop is somewhat slower (about 42.9s on two cores). The clock rate differnce is much less than the speed difference, so I think this says the 16-core machine has a better memory architecture than my laptop.  (What kind of machine did you run on?)

The C++ code is about 11% faster than the Cilk++ code on one processor. I'm still looking into that.