Collecting remote data in irregular problems with hierarchical representation of the domain

Abstract

Irregular problems require the computation of some properties of a set of elements irregularly distributed in a domain. These problems satisfy a locality property because the properties of an element e depend upon those of its neighbors according to a problem dependent stencil. This paper proposes two strategies, fault prevention and informed fault prevention, to collect properties of elements mapped onto remote processing nodes that minimize the corresponding overhead. We describe an MPI implementation of informed fault prevention and the experimental results in the case of the adaptive multigrid method.