Message relaying techniques for computational grids and their relations to fault tolerant message passing for the Grid
Résumé
In order to execute without modi cation Message Passing distributed applications on a computational grid, one has to address many issues. The rst to come is how let processes of two di erent clusters communicate. In this work, we study the performances of relaying techniques (passing messages to a middle-tier) to solve this issue. When using relays, messages and most of the nondeterministic behavior of nodes pass through the relays during the execution. This provides the ability to implement fault tolerance at the relay level using pessimistic message logging techniques.We also evaluate the overhead of this logging and study how relays should be designed and fault tolerance protocols composed to provide a full fault-tolerant Message Passing Interface library for computational grids.