This project proposes the development of a software tools for advanced
algorithms on a cluster of high performance graphics processing
units(GPUs). The initial goal of this library will focus on a single
driver application, the fundamental numerical study of Nuclear Forces
(Lattice Quantum Chromodynamics: LQCD), and a second target
application, the numerical simulation of the exciting nano-technology
of graphene. These are both multi-fermion problems well suited to
solution via multi-scale algorithms on many-core architectures. This
pilot project draws on experience gained by the small team at Boston
University and Harvard in developing Dirac solvers for lattice field
theory. Two building blocks from prior research are (1) the
construction of an adaptive multigrid (MG) solver for the Wilson Dirac
operator of LQCD, which demonstrates a 20x speedup compared to the
best Krylov solvers in production code and (2) a highly optimized Krylov
solver for the same operator on GPUs (but without multigrid),
realizing a 10x improvement in price/performance over traditional
clusters. The library will unite these feature and generalize their
domain of applicability.
As an example of the broader impact, it is estimated that combining
these two technologies (MG algorithms and GPU architectures) will yield
an 100-fold improvement in price/performance for the most
compute-intensive component of LQCD simulations. Such an advance
would be truly transformative, making an immediate impact in nuclear
and particle physics. At the same time, it will serve as a prototype
of the more generic problem of mapping hierarchical algorithms onto
heterogeneous architectures, a challenge of paramount importance on
the path to the exascale. The software library will be designed to bring
similar benefits to graphene technology and to evolve to accommodate
additional target application and additional domain decomposition
algorithm to mitigate the communication bottleneck of Exascale
designs.