Conveners
Algorithms and Machines
- Nigel Cundy
Algorithms and Machines
- Georg Bergner
Algorithms and Machines
- Hideo Matsufuru
-
Mr Greg McGlynn (Columbia University)14/07/2015, 14:00Algorithms and MachinesTalkWe discuss algorithmic improvements being used in the evolution of new 2+1+1 flavor lattices by the RBC/UKQCD collaboration. A technique similar to Hasenbusch mass splitting and suggested previously by Brower, Neff, and Orginos allows the light quark action to be split into multiple parts with different values of L_s, the size of the fifth dimension. This allows us to reduce chiral symmetry...Go to contribution page
-
Chulwoo Jung (Brookhaven National Laboratory)14/07/2015, 14:20Algorithms and MachinesTalkRecent advances in hardware, combined with various deflation techniques has made it crucial to optimize not only for solver performances, but also for memory and disk footprint to keep the overall efficiency high. Here we report on advances in Domain Wall Fermion formalism(DWF), such as zMobius which achieves a good approximation to the scaled Shamir DWF with smaller extent in 5th...Go to contribution page
-
Mr Andrea Bussone (CP3-Origins, University of Southern Denmark)14/07/2015, 14:40Algorithms and MachinesTalkImposing twisted boundary conditions on the fermion fields is a procedure extensively used when evaluating, for example, form factors on the lattice. Twisting is usually performed for one flavour and only in the valence, and this causes a breaking of unitarity. In this work we explore the possibility to restore the unitarity through the reweighting method. We evaluate the change in the...Go to contribution page
-
Dr Meifeng Lin (Brookhaven National Laboratory)14/07/2015, 15:00Algorithms and MachinesTalkThe application of the Dirac operator on a spinor field, the Dslash operation, is the most computation-intensive part of the lattice QCD simulations. It is often the key kernel to optimize to achieve maximum performance on various platforms. Here we report on a project to optimize the domain wall fermion Dirac operator in Columbia Physics System (CPS) using the R-Stream source-to-source...Go to contribution page
-
Prof. Peter Boyle (University of Edinburgh)14/07/2015, 15:20Algorithms and MachinesTalkWe discuss progress and performance from the development of a data parallel interface for cartesian fields of tensor data types. The library is suitable for the simulation of QCD including multiple grids. Wherever appropriate the interface is quite similar to that of QDP++, but the library makes use of C++11 features to reduce the volume of code, compared to QDP++ while bringing...Go to contribution page
-
Mr Matthias Rottmann (University of Wuppertal)16/07/2015, 10:40Algorithms and MachinesTalkIn this talk, we present a multigrid based eigensolver for computing low modes of the hermitian Wilson Dirac operator. For the non-hermitian case, multigrid methods have already replaced conventional Krylov subspace solvers in many lattice QCD computations. Since the gamma5 preserving aggregation based interpolation is valid for both, the hermitian and the non-hermitian case, inversions of...Go to contribution page
-
Mr Jakob Simeth (University of Regensburg)16/07/2015, 11:00Algorithms and MachinesTalkDisconnected contributions to hadronic quantities are noisy and still computationally challenging. Here, we explore the possibilities of Multigrid-based eigensolvers for Low-mode averaging and related methods. Using only the few lowest approximate eigenmodes of the Hermitian Dirac operator $\gamma_5 M$ allows us to avoid expensive exact solves and improve the signal efficiently. In this...Go to contribution page
-
Dr Hiroya Suno (RIKEN Advanced Institute for Computational Science; RIKEN Nishina Center)16/07/2015, 11:20Algorithms and MachinesTalkWe are developing a computer code for calculating eigenvalues of the non-Hermitian O(a)-improved Wilson-Dirac operator. We introduce here the Sakurai-Sugiura method, which is an eigensolver algorithm based on a contour integral, allowing us to calculate desired eigenvalues located inside a given contour. We report the test results for low-lying eigenvalues obtained with free-case, quenched and...Go to contribution page
-
Dr Alexei Strelchenko (FNAL)16/07/2015, 11:40Algorithms and MachinesTalkWe present a brief overview of deflation algorithms recently implemented in the SciDAC QUDA library for solving the lattice-Dirac equation with multiple right-hand sides on NVIDIA accelerators. In particular, we discuss implementation aspects of the mixed precision technique for the eigenvalue deflation which helps to considerably relax GPU memory requirements while allowing for an...Go to contribution page
-
Dr M Clark (NVIDIA)17/07/2015, 16:30Algorithms and MachinesTalkWe describe recent developments to the QUDA software framework, a library aimed at deploying lattice QCD computations on GPUs. The library has ever broadening support for various LQCD actions and algorithms, as well as being integrated into many LQCD applications. Recent focus has been on improving strong scaling for multi-GPU calculations and developing an adaptive multigrid solver. We...Go to contribution page
-
Dr hirokazu Kobayashi (Intel K. K.)17/07/2015, 16:50Algorithms and MachinesTalkWe implemented lattice QCD on Xeon Phi coprocessor using intrinsics as vectorization method, and OpenMP and MPI as parallelization method. Our implementation uses double precision conjugate gradient (CG) solver which also supports multi-shift CG. We present our optimization methodology and performance for key steps in CG algorithms.Go to contribution page
-
Dr Mario Schroeck (INFN - Roma Tre)17/07/2015, 17:10Algorithms and MachinesTalkWe present the implementation of twisted mass fermions within the QPhiX code package along with an interface to tmLQCD. We analyze the performance on Intel Xeon Phis and Intel Xeon Haswell CPUs. On the latter we find a speedup of larger than 4x for the mixed precision conjugate gradient inverter.Go to contribution page
-
Dr Hidekatsu Nemura (Center for Computational Sciences, University of Tsukuba)17/07/2015, 17:30Algorithms and MachinesTalkWe present our recent effort to develop the CUDA code to measure the Nambu-Bethe-Salpeter (NBS) wave function of various baryon-baryon systems on the lattice, which is a primary quantity to make a first-principle calculation of nuclear force and hyperonic-nuclear forces from lattice QCD. Not only the nucleon-nucleon (NN) but also the hyperon-nucleon (YN) and hyperon-hyperon (YY) interactions...Go to contribution page
-
Mr Jun Doi (IBM Research - Tokyo)17/07/2015, 17:50Algorithms and MachinesTalkAs supercomputers are shifting from peta-scale to exa-scale, computers with accelerators such as GPUs, MICs and FPGAs have become one of the big trends of supercomputer because of their low energy consumption and high density. Now IBM’s POWER processor has quite new power, Nvidia’s Tesla GPU brings huge computational capability. It is important for us to understand how this new POWER+GPU...Go to contribution page