Section: New Software and Platforms

Platforms

Fast linear system solvers in public domain libraries (http://icl.cs.utk.edu/magma/ )

Participant : Marc Baboulin [correspondant] .

Hybrid multicore+GPU architectures are becoming commonly used systems in high performance computing simulations. In this research, we develop linear algebra solvers where we split the computation over multicore and graphics processors, and use particular techniques to reduce the amount of pivoting and communication between the hybrid components. This results in efficient algorithms that take advantage of each computational unit [12] . Our research in randomized algorithms yields to several contributions to propose public domain libraries PLASMA and MAGMA in the area of fast linear system solvers for general and symmetric indefinite systems. These solvers minimize communication by removing the overhead due to pivoting in $LU$ and $LDLT$ factorization. Different approaches to reduce communication are compared in [2] .

See also the web page http://icl.cs.utk.edu/magma/ .

cTuning Framework (http://cTuning.org ): Repository and Tools for Collective Characterization and Optimization of Computing Systems

Participant : Grigori Fursin [correspondant] .

Designing, porting and optimizing applications for rapidly evolving computing systems is often complex, ad-hoc, repetitive, costly and error prone process due to an enormous number of available design and optimization choices combined with the complex interactions between all components. We attempt to solve this fundamental problem based on collective participation of users combined with empirical tuning and machine learning.

We developed cTuning framework that allows to continuously collect various knowledge about application characterization and optimization in the public repository at cTuning.org. With continuously increasing and systematized knowledge about behavior of computer systems, users should be able to obtain scientifically motivated advices about anomalies in the behavior of their applications and possible solutions to effectively balance performance and power consumption or other important characteristics.

Currently, we use cTuning repository to analyze and learn profitable optimizations for various programs, datasets and architectures using machine learning enabled compiler (MILEPOST GCC). Using collected knowledge, we can quickly suggest better optimizations for a previously unseen programs based on their semantic or dynamic features [10] .

We believe that such approach will be vital for developing efficient Exascale computing systems. We are currently developing the new extensible cTuning2 framework for automatic performance and power tuning of HPC applications.

For more information, see the web page http://cTuning.org .

NT2 (http://www.github.com/MetaScale/nt2 )

Participants : Pierre Esterie, Joël Falcou, Mathias Gaunard, Ian Masliah, Antoine Tran Tan.

NT2 is a C++ high level framework for scientific computing.[18]

Boost.SIMD (http://www.github.com/MetaScale/nt2 )

Participants : Pierre Esterie, Joël Falcou, Mathias Gaunard.

Boost.SIMD provides a portable way to vectorize computation on Altivec, SSE or AVX while providing a generic way to extend the set of supported functions and hardwares.