Section: New Results

HPC Component Model

Participants : Zhengxiong Hou, Vincent Pichon, Christian Pérez.

L2C: A Low Level Component Model

We have proposed a low level component model (L${}^2$C) that supports directly native connectors for typical scenarios of high performance computing, such as MPI, shared memory and method invocation [10] . We have applied it to a typical example of stencil computation, i.e., a 2-D Jacobi application with domain decomposition. The experimental results have shown that L${}^2$C can achieve the equivalent performance as native implementations, while gaining benefits such as performance portability on the basis of the software component model.

Auto-tuning of Stencil Based Applications

We started modeling the performance of stencil applications on multi-core clusters. We focused in particular on a 2D Jacobi benchmark application and the NEMO application as well as memory bandwidth performance. We derived a tuning approach including data partitioning within one node, the selection of the number of threads within a multi-core node, a data partitioning for multi nodes, and the number of nodes for a multi-core cluster. This model is based on a set of experiments on machines of Grid'5000 and on Curie and Juqueen supercomputers. A paper presenting these results is in preparation.