How to increase feed-forward NN training performance?

Hello,

General info about the system:

Epoch: 2000

Hidden Layer: 300

Training Function: trainbr

Input size: 3*17490

Output size: 1*17490

(Actually values are aound 150K but right now I'm using a filters)

1-) When the training should stop according the parameters?

2-)I've tried many training and all stopped at the end of epoch number. Is it ok to stop at the end of epoch value?

3-)What the importance of performance value? According to performance value point, which values can be acceptable? I mean is there any acceptable range?