local/distributed neural network performance questions

when matlab trains a large sample size (say with >1M), and a good size nn (say, 400 input features, 500-1000 hidden units on 2 layers, and 2 output classes) on local or distributed nodes:

1) does it use mini-batch during training?

2) how does it decide size of mini-batch size?

3) if I run it on multiple distributed nodes, I assume the mini-batch size can be specified within each data set I specify for each distributed nodes?

Thanks, Hc