The development of capable exascale systems was made possible by an interdisciplinary and collaborative co-design process. The Exascale Computing Project (ECP) created co-design centers such as the Center for Online Data Analysis and Reduction (CODAR) to develop algorithms and software components optimized for cutting-edge hardware. The centers’ objective is to advance computational motifs common to multiple application development projects, thereby fostering participatory development processes to meet the complex and often conflicting needs of current and future exascale applications. The co-design teams worked closely with application developers to deliver efficient and reliable software products that are integral to the unprecedented results generated on exascale supercomputers such as Frontier.
The world’s fastest supercomputers perform 1018 operations per second, but write information to disk at only 1012 bytes per second. The gap between operations and outputs is rapidly growing larger as computational technology advances—for example, the first petascale supercomputers had a compute to output ratio 200 times smaller than modern exascale supercomputers. In this new world, applications must increasingly perform online data analysis and reduction to make full use of new computational hardware. Data analysis and reduction tasks introduce algorithmic, implementation, and programming model challenges that are unfamiliar to many scientists. These challenges have significant implications for the design of exascale software systems, and can slow the pace of scientific discovery on exascale machines without adequate expertise and support.
The ECP established CODAR to produce computational infrastructure for online data analysis and reduction on the world’s fastest supercomputers. The co-design center was critical in developing and implementing services to reduce the amount of data that must be written to disks for offline analysis without significantly impacting data quality. The CODAR team collaborated with application developers to create software libraries that are essential to the functionality, accuracy, and efficiency of many exascale applications. CODAR also facilitated collaborations to release software artifacts and construct application-oriented case studies to further support the functionality of exascale applications.
To enable online data analysis and reduction, the CODAR team had to create software capable of pairing these processes with on-node computations during an ongoing simulation. The team targeted common data analysis and reduction methods—such as compression and feature and outlier detection—as well as methods specific to particular experimental domains, then designed and implemented new methods that can be run in situ. CODAR also created tools to measure the performance tradeoffs between online and offline data analysis in application-specific contexts. These tools guide application development and empower performance and precision.
The online data analysis and reduction methods developed by CODAR will be critical for efficiently utilizing computational power as the output to compute ratio of high-performance computers continues to expand. The co-design team has already used these methods to significantly improve the performance of several ECP applications, develop an interactive visual analytics system for monitoring and altering data compression, and implement new machine learning solutions for data analysis and reduction. The CODAR software suite will also serve as a foundation for future efforts to integrate multiple simulations and analyses onto a single computational system for optimized performance and accelerated scientific progress.