Storage Systems and Input/Output for Extreme Scale Science

The processing capability of computing systems, especially supercomputers, continues to increase rapidly. Within 10 years, exascale platforms will increase computational performance by a factor of at least100 compared to 2015’s 10-petaflop systems and they will support billionway concurrency. However, improvement in the performance and capacity of storage systems and input/output (SSIO) bandwidth have lagged considerably, so that available storage and input/output bandwidth on exascale platforms will increase by only a factor of three compared to current technology. Improvements in storage efficiency will be achieved in part by increasing the complexity of the memory hierarchy to include non-volatile random access memory (NVRAM), which significantly complicates use of the system by scientists.

These trends combine to force significant changes in the workflow for computational science, which must shift from saving data for post-hoc analysis to incorporating various forms of data analysis and visualization during the run of a simulation, with comparatively little data saved for post-hoc analysis. These workflow changes require significant modifications to the way data are accessed and manipulated on the supercomputer, as well as how data are stored. Other science user facilities are similarly affected by SSIO limitations, with increasingly severe data reduction required for experimental/observational research projects.

Furthermore, as we look towards the extreme scale era, storage systems will become so large in numbers of devices/parts that we can no longer assume that existing designs of system components can be counted on to always be available and never lose/damage data. New designs are needed to meet extreme scale science requirements for usability, scalability and reliability. 2 To address these concerns, the Office of Advanced Scientific Computing Research (ASCR) invites computer science research proposals to address three themes:

  1. Measurement and Understanding of Storage Systems and Input/Output Challenges;
  2. Scalable Storage Software Infrastructure; and
  3. New Paradigms in Storage Systems and Input/Output


Modeling Impacts of Resilience Architectures for Extreme-Scale Storage Systems
A Framework for Holistic I/O Workload Characterization
A Software Defined Storage Approach to Exascale Storage Services
Proactive Data Containers for Scientific Storage
SIRIUS: Science-driven Data Management for Multi-tiered Storage
UNITY: Unified Memory and Storage Space