Weka is probably the most popular general purpose machine learning library. But it can be quite slow in my experience.
I have been looking at Shark, Waffles, dlib, Plearn, and MLC++ as alternatives. Of these, Shark and dlib look the most promising.
Does anyone have any experience when it comes to performance testing of these libraries?
For me, what matters most would be "Does this toolkit have the algorithm or feature I want to try out?" Since these toolkits provide a fairly diverse set of features you should first try to narrow down what it is you want to do.
So for example, if you have a burning desire to try out different evolutionary optimization algorithms then I would go with something like Shark.
On the other hand, I prefer dlib for most of my work but that doesn't necessarily mean a lot since I wrote it :) However, if you are interested in binary classification then let me suggest my current favorite method for that, the svm_c_ekm_trainer. I frequently use this to train non-linear SVMs on datasets of hundreds of thousands of points. It usually runs in a few minutes (or sometimes even seconds) while the classic SMO algorithm for this would take hours or days to finish.
There were also some good answers to a similar question asked not too long ago: Which machine learning library to use.