Tanszékcsoporti szeminárium

Distributed aggregation methods, such as consensus or gossip-based approaches, have been studied intensely in the last decade and many fundamental contributions have been made. These concepts primarily target loosely coupled distributed systems,
such as sensor networks or P2P systems, and have so far been hardly considered in the context of high performance computing. Moreover, to the best of our knowledge,
some issues arising in practice and algorithmic strategies for recovering from silent failures (maintaining convergence to the correct limit!) still need to be addressed in more detail.

We discuss a general approach for provably reliable distributed aggregation under more realistic assumptions on the distributed environment. By taking theoretical as well as practical points of view into account, we point out some shortcomings
in existing methods and their analyses. We derive reliable distributed
aggregation algorithms which provably tolerate silent message loss. For proving their correctness and convergence, we need to extend the existing convergence theory for distributed aggregation algorithms. Beyond that, we also discuss some challenges beyond silent message loss which are of particular interest in a high performance
computing context.