PhD students' day — CITI lab


Satisfying Dataflow Programs Constraints on Multicore Architectures Through Runtime Adaptations

by Manuel Selva

Download slides
Picture of Manuel Selva

Details

Socrate team
3rd year

Keywords

Dataflow
Multicores
Compilation
Runtime Adaptation
NUMA
Througput

Advisors

Lionel Morel
Stéphane Frénot

Partner

Bull


In the context of multi-core processors and the trend toward many-core, dataflow programming can be used as a solution to the parallelization problem. By decoupling computation from communication, this paradigm naturally exposes parallelism in several ways. Targeting common multicores architectures, dataflow compilers first translates programs into multi threaded applications in a low level language such a C or C++ and then use standard compilation tools to produce binary code.

We propose language extensions for expressing throughput properties over dataflow programs. These properties are taken into account into the compilation tool chain. We introduce run-time mechanisms for the observation of the effective throughput and the identification of bottlenecks if this effective throughput doesn't match the one expressed at source program level. For the identification of bottlenecks, we leverage modern hardware performance monitoring units.


Acknowledgements

CITI lab

INSA Lyon

Inria

HiKoB

INSA Valor

Rhône Alpes region