Supertwin: Digital twins for high-performance computing clusters

Official URL: https://risc01.sabanciuniv.edu/record=b2993554_(Table of contents)

Computational systems are extremely complex and the composition of their hardware and software components greatly vary from machine to machine. This nonstandardized environment can cause up to 100% difference between the best and worst completion times with the same input data. On top of that, the shape of the input data and executed kernels add even more variance to the situation. However, computational systems are not completely hostile environments. These systems are also equipped with diverse observability capabilities. A typical Linux system can report thousands of real-time execution and performance-related metrics from both its hardware and software components. Digital Twins are knowledge management systems that have vast application areas in the industry, however, digital twins of computational systems remain a gap in the literature. SuperTwin is a knowledge representation generator and manager of the tools and performance data that interact with it. It creates a digital twin of a computational system via detailed probing, configures and listens to performance metric samplers, creates real-time visualizations, links the acquired information, and enables semantic queries for advanced analysis. In this work, design and implementation choices for SuperTwin are thoroughly presented. The effect of profiling on remote systems is analyzed and the accuracy of the readings is investigated.