sampling with replacement from matrix with clustered data

Hello,

I'm looking for a clever (computationally efficient) way to sample from a data matrix in clusters.

Given matrix A(n,I) where n is large, and I is small, it's easy and fast randomly resample from A by:

B=A(ceil(rand(n,1)*n),:).

But, lets say that A(:,10) and A(:,11) are indexes that represent two clusters. That is for each of the clusters in A(:,10), A(:,11) is an index of the second cluster. To give specifics, A(:,10) is four clusters and within these four clusters there are many smaller clusters (also of 4). But because n is large (about 200,000) within each of the first clusters, there are many small clusters.

In practice, what I want to is create a matrix that has n rows and is made by drawing with replacement from the first cluster and then the second cluster.

I have done this in steps by drawing from the first variable and then the second and pulling out the rows in the set (which turns out to be 4 rows) and building up a matrix (by appending each draw to the previous matrix) and doinng this many times in a for loop. It takes about 5 minutes too create one of these resampling matrices, which is too long.

Can someone point me to a clever routine or approach to improving this?

Many thanks!