Thanks! I will do some follow-up posts on performance, but know that it has been a MAJOR design consideration.
Consider the following in Saddle:
val s1 = Series(vec.rand(10000), Index(Vec(array.randIntPos(10000)) % 100))
val s2 = Series(vec.rand(10000), Index(Vec(array.randIntPos(10000)) % 100))
clock { s1.join(s2, how=index.OuterJoin) }
This clocks in at 19ms on my machine after Hotspot kicks in.The equivalent pandas:
In [10]: ix1 = np.random.random_integers(0, 100, 10000)
In [11]: ix2 = np.random.random_integers(0, 100, 10000)
In [12]: df1 = DataFrame({'x' : np.random.rand(10000)}, ix1)
In [13]: df2 = DataFrame({'y' : np.random.rand(10000)}, ix2)
In [14]: %timeit df1.join(df2, how='outer')
10 loops, best of 3: 37.7 ms per loopPS Regarding EJML, after extensive research, I found it hands down the fastest pure-java implementation for doing linear algebra.
While it's maybe 2x-4x slower than JNI wrapped ATLAS or MKL, for the cases I deal with, it just doesn't matter vs ease of use.
That said, it's LGPL, so I made it easy to swap out for other matrix libraries if you need.
Do we have performance information yet, even on some basic, common use cases?
Also, the docs mention EJML as the backend for Saddle's data structures--do you have any thoughts on using EJML?