Section: New Results

Process Networks with routing for parallel architectures

Participants : Robert de Simone, Emilien Kofman, Jean-Vivien Millo.

In the past we developed a dedicated Process Network (PN) formalism with explicit static switching/routing schemes for data flow. This year we considered the practical use of our formalism to model data-streams in specific applicative contexts.

In a first direction we considered the case of stencil algorithms, usually modeled with cellular automata (CA) (as in heat or gas propagation models for instance). In that case, the application itself is modeled in a way strongly similar to a physical architecture consisting of a regular mesh/array of parallel processors (MPPA). Mapping can seem to be straighforward then, safe that the neighborhood and connection topology may differ from the CA model to the MPPA. Our results consider efficient routing and propagation schemes on a given MPPA interconnect fabric, so as to match all-to-all broadcast paterns up to a given distance (on the CA topology). They are described in [20] , and were implemented on Kalray MPPA256 prototype architecture.

A similar modeling effort was conducted, this time on FFT algorithm models (again described as parallel pipe-lined tasks). Again switching/routing schemes were provided in our formal PN model to map the virtual logical dependences onto concrete connection patterns in a MPPA256 model. This was the subject of Emilien Kofman internship, of which preliminary results were presented in a junior workshop [36] .