number of hidden layer nodes in neural network fitting function

In am using neural network fitting function with four input neuron one hidden layer/the matlab default/ and one output neuron for prediction. I start using 10 hidden layer neurons.
1. My problem is how can measure the performance and when to stop increasing the hidden layer nodes?
2. In training the network; when shall I stop, am not clear on the performance metrics.
3. How can I test the nnfit for new data?