Scientific applications have been using decades old interfaces and complex I/O software layers. These layers in the parallel I/O software stack have strict consistency semantics and interdependent performance characteristics that leave the burden of performance tuning on application developers. We are now in the exascale era; however, the storage and I/O software pose familiar problems with performance. With the trends of heterogeneity of computing devices, application landscape with AI and experimental / observational data analyses, and deep and distributed memory and storage hierarchies, we need to rethink the designs of data storage and I/O on HPC systems. In this talk, I will present a few solutions we developed to cope with scalable I/O using existing libraries and then present a vision for future storage and I/O systems aiming automatic data management.