We can use machine learning/statistics in plugins to generate more 'real; disk images/extents. Currently, sparse-random is the go-to test. I think we can leverage ML/statistics for more true-to-life disk images hence more accurate benchmarks. These will expose factors that are more relevant for real systems. Generative Adversarial Networks(GANs) might be useful. However, we will need info about real images to first train the generator. Regards Abhay