Description:black-box analysis, engineering modeling
Abstract:Large and complex chip model testing is computationally intensive task. It is not possible to test every single input and get output correlations for all tests. Search space is too big to fit any supercomputer. However, using Monte Carlo methods and splitting computations across grid it is still possible to accumulate enough information about black-box inner workings. Grid resources were allocated and utilized to achieve superior results.
1. Framework for black-box analysis
2. Autonomous tracking of computations
3. Search space visualization
Fourty-seven black-box chip models, some with at least few hundred inputs and outputs, were analyzed using grid. Computations were split in a parallel matter to cover as much test sets as possible. Single submitted tasks had only one restriction: time limit. Many attempts to finish tasks in a specified time frame had failed before due to long queues. Tasks would get scheduled and were not able to use all proxy time requested by user proxy. In many cases tasks spent more time being Scheduled than Running. So another, more flexible restriction was substituted. Time limit was set only when task started running on the Working Node. Time left for computation was then derived from voms proxy information and CPU time limit. In such way it was possible to get the most efficient use of resources to complete model analysis.
Programming languages: C++, Perl.