Assuming that there are p processors, all roughly identical, the sample selection procedure at Layer 1 divides the data up approximately evenly amongst them.
If we focus on landscape-style data, that is achieved by giving each processor all the instances for 1/p th of the total number of attributes.