Speed is logarithmic.
20% is logarithmic-speak is "tiny".
20% is logarithmic-speak is "tiny".
You draw those X and Y according to them.
I remain skeptical of these micro benchmarks. Almost every time I've messed around with porting pieces of C++ to Java or C# and benchmarking it I get a large difference in favor of C++.
And the fastest way to use it this way is Java, which is what the number crunching users of Hadoop use. Which is to say that the emerging big data compute platform is Java. Which is to say that you're wrong.
Why would Java be the fastest way to do computations with Hadoop? You can use many languages for the map/reduce pieces. The fastest language would be fastest language, not necessarily Java.
As to why streaming is slow: I don't know. It just is. Check it out. C++ streaming performance is the pits.