xgboost histogram can't scale to 64 threads (scales negatively?)

[Laurae2]

The benchmark on this page shows xgboost histogram on Bosch, but it is slower than a single thread of i7-7700HQ throttled to 2.8 GHz (which trains with even more data).

Number of threads should be set to a low number (such as 8) and not a huge number (p3.16xlarge uses 64 threads), it is probably a negative scaling.

[RAMitchell]

I was hesitant to set any parameters away from their default values but I think I will do this for the next run.