We present an algorithm to dynamically adjust the data assigned for each worker at every epoch during the training in a heterogeneous cluster. We empirically evaluate the performance of the dynamic partitioning by training deep neural networks on the CIFAR10 dataset.