As HPC resources evolve into petascale resources and beyond, substantial mismatches in scale between the computation resources and the storage resources demand rethinking how to manage generated data for scientific discoveries. Process counts of 100,000s, 1,000,000 or more overwhelm storage resources causing IO to consume too large a percentage of total runtime. Shared scratch file systems that facilitate end-to-end processing by using multiple HPC resources compound the problem as much as they help. While it is tempting to perform micro optimizations to aid either writing, reading, or some analysis tasks, only optimizations that address the entire end-to-end science process will ultimately be useful. By carefully managing data output techniques mindful of later data analysis tasks, both write and read performance can be improved. Further, by incorporating `in transit' processing, data generation runtimes can be reduced even when considering the additional resources employed while adjusting the data to be better annotated, filtered, or processed aiding the analysis scientific discovery process.