CyQuarks

PI: Giannis Koutsou

Institutions: The Cyprus Institute (CYI), University of Cyprus (UCY), Deutsches Elektronen-Synchrotron (DESY), Istituto Nazionale di Fisica Nucleare (INFN)

Application Area: The application is in the domain of lattice Quantum Chromodynamics (lattice QCD), which is a method for solving numerically QCD, the fundamental field theory which describes the interactions between quarks, binding to form hadrons such as protons and neutrons.

Goal at EuroHack: Port deflation (using ARPACK) to GPUs.

Challenges: Many sparse matrix times dense matrix multiplications; many kernels had to be implemented as macros

Results: Roughly 50% speedup w.r.t. 1 Sandybridge socket 8 OMP threads

Experiences: refactoring was needed (some changes unpleasant); some optimizations difficult to maintain; hackathon was an excellent opportunity, better than tutorials with toy examples; ease-in day would have been useful (Sunday?)

Future: CUDA code will be used in production, will make the initial setup much faster.

Final Presentation