RadFlow: An Interest-Centric Task Based Dataflow Runtime

Abstract

We present RadFlow a runtime system for task based Dataflow parallel application using an interest-centric network protocol to data communication among task. The RadNet protocol ability to decouple data destinations from its node IP addresses allows Rad Flow more flexibility, enabling mechanisms like computation migration and elastic tasks to be carried out. We also demonstrate how to create a Bag-of-Task, a fork/join, as well as an elastic fork/join Dataflow parallel application for the Rad Flow runtime. Furthermore, an elastic Dataflow application provides the application developer means to cope with the failures rates in future Exascale environments.