Franck Cappello, a senior computer scientist is Argonne’s Mathematics and Computer Science Division, was a plenary speaker at CoDA2016. The meeting, held March 2–4 in Santa Fe, aimed at exploring data-focused research across the U.S Department of Energy (DOE).
The invited program included six themed sessions, each with a plenary speaker. Cappello was the plenary speaker for the session on cybersecurity. His presentation focused on a major concern facing extreme-scale computing: trust in results of numerical simulations and scientific data analysis.
Cappello first spent a fair amount of time giving examples of errors, bugs, and attacks leading to corruptions in results. He also discussed current ways of detecting and improving trust. One popular technique, V&V, involves validation, which determines the faithfulness of the computational model to the real world, and verification, which determines the faithfulness of the code to the numerical model. Many of the problem classes for which V&V methods exist, however, do not overlap with the complex simulations run at DOE. Uncertainty quantification (UQ) is another technique used to improve the trust in simulations and data; the aim here is to account for all sources of uncertainty introduced by the mathematical model. UQ may not, however, detect errors arising from spurious software, for example.
“Since the trust problem encompasses all levels of the stack – from the hardware and operating and runtime system to numerical libraries and applications – and is related to many aspects of numerical simulation and data analytics, innovative detection techniques clearly are needed, as well as codesign research to explore the best way to implement trust mechanisms,” said Cappello.
To this end, he presented a promising new approach: the external algorithmic observer. The main idea is that the external observer verifies that the observed execution respects constraints set by the application developer or user. Cappello presented several implementations of this approach and discussed the tradeoffs between performance overhead and detection accuracy.
For information on the CODA2016 meeting, see the website.
For further information on trust in numerical simulations and data analysis, see the white paper by Cappello and his colleagues.