Grid result checking

• Authors: Cécile Germain-Renaud;Dephine Monnier-Ragaigne
• Venue: Proceedings of the 2nd conference on Computing frontiers
• Year: 2005

Result checking is the theory and practice of proving that the result of an execution of a program on an input is correct. Result checking has most often been envisioned in the framework of program testing or property testing, where the issue is the conformity of the program to some a-priori specification. Very large scale distributed computing systems demand to tackle the issue of computation correctness, albeit from hypothesis very different from the program testing ones. The general issues examined in this paper are the following. First, the definition of checking methods adapted to large-scale Monte-Carlo simulations; for these applications, no external criterion can be used to assess the quality of the result. Second, two result checking algorithms which minimize the overall overhead through an adaptive strategy. Finally, a specialization of this framework to a case study, the Auger astrophysics experiment. Our main contributions are: first to focus on checking Monte-Carlo simulations, which have rarely been considered previously; second to define a probabilistic checking strategy including the risk of first kind (false positive) as well as the risk of second kind (false negative) which is usually the only one considered, and which is compatible with Byzantine saboteurs; third, to exploit the probable characteristics of the behaviour of the saboteurs to optimise for the most frequent case. Finally, we show on a case study that the implementation details can be carried out