How to Train a Model on a Cheap Cluster Using Block Coordinate Descent