Section: New Results
Enabling high performance applications on emerging architectures
Participants : François Bodin, Damien Fétis, Junjie Lai, André Seznec.
To achieve very high performance in some application domains, architectures must be specialized. We are involved in an ANR project aiming at defining an architecture for Lattice QCD (Quantum ChromoDynamics) and a “Pôle de compétitivité” project aiming at defining a powerful platform for embedded systems.
Architecture for Lattice QCD
Participants : François Bodin, Junjie Lai, André Seznec.
Simulation of Lattice QCD is a challenging computational problem that requires very high performance exceeding sustained Petaflops/s. In the framework of the ANR Cosinus PetaQCD project, we will model the demands of this application on the memory system and synchronization mechanisms. The objective is to obtain a first order comparison of different design options for a LQCD machine based on off-the-shelf multi-cores or multi-cores+accelerators designs, therefore guiding the dimensioning of a dedicated machine for LQCD. The methodology should be able to be adapted to the study of other massively parallel applications to understand their performance behavior. It should also be useful in early multi-core design phases to help to decide on internal die organization such as number of cores vs cache size, hierarchical organization, etc.
Data Locality Analysis of Parallel C Programs
Participants : François Bodin, Guillaume Papauré.
In the context of the POPS project, we have studied strategies to achieve efficient parallel execution based on OpenMP and assuming asymmetric core behaviors [36] . This work aims at understanding how to limit performance degradation of OpenMP programs when running on computation nodes that are shared by multiple applications. This competition for hardware resources is a particularly important efficiency factor in the context of hybrid parallel programming (e.g. mix of OpenMP and MPI) or in application servers used in a multi-user context.
Backend optimizations: Sofan for the Ter@ops project
Participants : François Bodin, Damien Fétis.
In the context of the ApeNEXT project [1] , we have developed over the last few years SOFAN (Software Optimizer for ApeNEXT). This software optimizer attempts to explore different back end optimization strategies for ApeNEXT applications. In 2007, a production version of SOFAN was delivered to the users of ApeNEXT. In the context of the Ter@ops project, SOFAN has been retargeted towards one of the accelerators of the Ter@ops machine, the Thomson FIRE EVO coprocessor. FIRE EVO is SIMD coprocessor with an array of 16 VLIW processing units. A gcc code generator was developed by the Alchemy EPI for this accelerator. SOFAN was adapted to realise the control flow and data flow analysis of SIMD assembly code. It also optimizes the VLIW Processing Units resources utilisation.
