TY  - CONF
AU  - Di Napoli, Edoardo
TI  - The ChASE library on distributed and heterogeneous platforms
M1  - FZJ-2016-04275
PY  - 2016
AB  - We  propose  to  step  away  from  the  black-box  approach  and  allow  the eigensolver to accept as much information as it is available from the application.  Such a strategy implies that the resulting library is tailored to the specific application, or class of applications, and loose generality of usage. On the other hand, the resulting eigensolver maximally exploits knowledge from the application and become very efficient.  With this general strategy in  mind,  we  present  here  a  version  of a Chebyshev Accelerated Subspace iteration Eigensolver (ChASE) which targets extremal eigenpairs of dense eigenproblems.  In particular, ChASE focuses of on a class of applications resulting in having to solve sequences of eigenvalue problems where adjacent problems possess a certain degree of correlation.  A typical example of such applications is Density Functional Theory where the solution to a non-linear partial differential equation is worked out by generating and solving dozens of algebraic eigenvalue problems in a self- consistent fashion over dozens of iterations. Similarly, any non-linear eigenvalue problem, which can be solved by the method of successive linearization, gives rise to sequences of correlated algebraic eigenproblems that are the target of ChASE. We re-design the eigensolver so as to minimize its complexity and have better control of its numerical features.  Following the algorithm optimizations, we strive to adopt  a  strategy  leading  to  an  implementation  that  would  lends  itself  to high-performance parallel computing and avoid, at the same time, issues related to portability to heterogeneous architectures.  We achieve such a goal  by  implementing  parallel  kernels  for  the  modular  tasks  of  the  eigensolver using programming strategies out of MPI, OpenMP, and CUDA.
T2  - Parallel Matrix Algorithms and Applications
CY  - 6 Jul 2016 - 8 Jul 2016, Bordeaux (France)
Y2  - 6 Jul 2016 - 8 Jul 2016
M2  - Bordeaux, France
LB  - PUB:(DE-HGF)6
UR  - https://juser.fz-juelich.de/record/811971
ER  -