Skip to content

Computational/Applied Math Seminar: Algorithm Based Solver Resilience

Wednesday, 23 October, 2013

SPEAKER:  Miroslav Stoyanov, ORNL

TITLE:  Algorithm Based Solver Resilience with Respect to Silent Hardware Faults

ABSTRACT:  The exponential increase of computing power over the past few decades has lead to a corresponding decrease in hardware reliability. At the same time, numerical algorithms have been developed under the assumption that all computations are performed accurately. As a result, modern solvers are very susceptible to a particular class of silent errors that may modify the result of a numerical operation without any external indication. The ever increasing rate of silent faults has presented a new and unanticipated challenge.

We extend the current framework of numerical analysis by removing the assumption that all arithmetic operations can be computed accurately to machine precision. We introduce the concept of ``hardware error'' added to our numerical approximations by potentially unreliable hardware. Using rigorous analysis, we develop new algorithms that minimize the propagation of hardware error and therefore guarantee convergence even when faults are encountered.

Venue

Ayres Hall
Room 122
1403 Circle Drive
Knoxville, TN 37996

Event Contact

Phone: 974-2463

The flagship campus of the University of Tennessee System and partner in the Tennessee Transfer Pathway.