ADIOS is a high performance publish/subscribe I/O framework which has been designed and developed for the exascale computing era. ADIOS provides an I/O abstraction unifying storage and staging I/O, a data model, self-describing data containers and a flexible framework that implements specialized solutions for different storage and network media. In-memory coupling of codes with subsequent data analytics platforms that allow for streaming analysis, in-situ visualization and combining AI and simulation, are now easier to build, to scale and to port to new systems using ADIOS for I/O at all levels of data movement in the workflow. Most of the day will be spent to teach all the data transfer capabilities in ADIOS-2, writing and reading files and in situ. Since reading is very important to all users, we will take extra time to show reading in python scripts beside the C++ analysis code in our workflow. In a snapshot, ADIOS:
- Is integrated into most of the popular analysis and visualization packages.
- Has strict continuous integration practices, providing stable, portable and efficient I/O services.
- Has a programming interface designed for easy switching from files to streams (on-HPC machines) to streams over the Wide Area Network.
On the first day of this tutorial, Dr. Scott Klasky and Dr. Norbert Podhorszki will introduce the ADIOS I/O framework and describe how it is being used, including: 1. In data intensive supercomputing simulations, using it for data storage and retrieval moving petabytes of data in a single job, 2. In coupled simulations using it for data movement between the simulation codes and to in situ analysis and in situ visualization services, 3. In AI applications, for collecting training data from ensembles of computations on the fly, and 4. In collaborations between experimental facilities and HPC centers to stream experimental data for near-real-time decision making.
Students are encouraged to bring with them their I/O challenges for discussion.