Section: New Results
HPC Component Model
Participants : Zhengxiong Hou, Vincent Lanore, Christian Perez.
Auto-tuning of Stencil Based Applications
We have finished designing a tuning approach for stencil applications on multi-core clusters [25] . We focused in particular on a 2D Jacobi benchmark application as well as memory bandwidth performance. The tuning approach includes 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 the Curie supercomputer.
Static 2D FFT Adaptation through a Component Model based on Charm++
Adaptation algorithms for HPC applications can improve performance but their implementation is often costly in terms of development and maintenance. Component models such as Gluon++, which is built on top of Charm++, propose to separate the business code, encapsulated in components, and the application structure, expressed through a component assembly. Adaptation of component-based HPC applications can be achieved through the optimization of the assembly. We have studied such an approach with the adaptation to network topology and data size of a Gluon++ 2D FFT application. Preliminary experimental results obtained on the Grid'5000 platform show the suitability of the proposed approach.
Towards Scalable Reconfiguration in Component Models
Some HPC applications require reconfiguration of their architecture at runtime; examples include adapting to (cloud) resource elasticity, efficient distributed deployment, Adaptive Mesh Refinement (AMR), and load balancing. This class of applications raises challenges such as handling of concurrent reconfigurations and distributed architecture representation at runtime. To our knowledge, no existing programming model addresses those challenges in the general case with both high programmability and scalability. We have identified a list of specific subproblems and use-cases and we have devised a preliminary component model to address some of them.