We're lucky in the sense that the IO for LQCD is small (compared to other scientific applications), in that we're usually only reading or writing gigabytes to terbytes. But also our code uses parallel HDF5 and it's someone else's job to make sure that works well :)