Thanks @Joseph Kambourakis​ ! It seems we could do the distributed XGBoost training using the num_workers regards to how many workers in the cluster. But can we also speed up by setting a parameter utilizing the number of cores in the cluster?