Picking step size for gradient descent

I'm writing a least-squares linear regression program. (as a learning exercise) It mostly works, but I can't figure out how to select the initial step size when doing the gradient descent. What are good techniques for that?