Towards Abinit on ExaScale supercomputers : the challenge for electronic structure physicists.

In Abinit, physical properties are computed from the state-of-the art first principles methods.
With several levels of parallelism with MPI, one cannot offer to rewrite the million of lines of the code each time supercomputer architectures evolve.
To avoid obsolescence and ensure the usability of the code, we introduced an abstract layer for intensive computing tasks which is be able to adapt to the hardware.
We started this work with CPU and Intel Xeon Phi.
We introduced an hybrid MPI/OpenMP parallelization to better suite on Intel Xeon Phi and many cores architectures.
We improved memory management, and destress memory bandwidth while increasing the FLOPS on Intel Xeon Phi.
Performance measurements show that with DDR and high speed MCDRAM memory, the ported part of the code runs similarly.
And the hybrid parallelization we introduced, allows us to perform new extra large calculations previously out of range.
Finally, with new cpu capabilities, it is crucial to notice the effect of vectorization or more generally the out of order numerical operations which increase numerical noise and favor instabilities.
We tackle the problem with stochastic method to identify parts of code that need attention.
All the problems we are facing will be even more critical in the area of exascale computing.