Ronald Babich
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Ronald Babich has a total of 16 co-authors. Here are the top co-authors by number of shared publications:
- Richard C. Brower
Boston University
7 shared publications - Claudio Rebbi
Boston University
5 shared publications - Michael A. Clark
Initiative in Innovative Computing 29 Oxford Street 02138 Cambridge MA
4 shared publications - Bálint Joó
Thomas Jefferson National Accelerator Facility Newport News 23606 VA USA
3 shared publications - Michael A. Clark
Center for Computational Science Boston University 3 Cummington Street 02215 Boston MA USA
3 shared publications - George T. Fleming
Fermi National Accelerator Laboratory
2 shared publications
Latest Publications
- Open Access
- 5 July 2012
- Published by Sissa Medialab Srl
- Vol. 139, 033
Progress on the QUDA code suite
At the time of Lattice 2010, we were about to announce a distribution of the code (QUDA 0.3) that supported both Wilson/clover and improved staggered quarks for computation on a single GPU. Multi-GPU code was running for both solvers, but with the restriction of grid partitioning in only the time dimension. In the past year, we developed code that allows us to cut the lattice in all four dimensions. This allows us to scale computations to order 100 GPUs yielding multi-teraflop performance. We will present results for both types of solvers on GPU clusters and for other kernels important for physics projects. We also compare performance and cost-effectiveness of full application codes running on CPUs with our GPU accelerated code.
- 5 July 2012
- Published by Sissa Medialab Srl
- Vol. 139, 033
Open AccessProgress on the QUDA code suite
At the time of Lattice 2010, we were about to announce a distribution of the code (QUDA 0.3) that supported both Wilson/clover and improved staggered quarks for computation on a single GPU. Multi-GPU code was running for both solvers, but with the restriction of grid partitioning in only the time dimension. In the past year, we developed code that allows us to cut the lattice in all four dimensions. This allows us to scale computations to order 100 GPUs yielding multi-teraflop performance. We will present results for both types of solvers on GPU clusters and for other kernels important for physics projects. We also compare performance and cost-effectiveness of full application codes running on CPUs with our GPU accelerated code.
- Open Access
Scaling Lattice QCD beyond 100 GPUs
Over the past five years, graphics processing units (GPUs) have had a transformational effect on numerical lattice quantum chromodynamics (LQCD) calculations in nuclear and particle physics. While GPUs have been applied with great success to the post-Monte Carlo "analysis" phase which accounts for a substantial fraction of the workload in a typical LQCD calculation, the initial Monte Carlo "gauge field generation" phase requires capability-level supercomputing, corresponding to O(100) GPUs or more. Such strong scaling has not been previously achieved. In this contribution, we demonstrate that using a multi-dimensional parallelization strategy and a domain-decomposed preconditioner allows us to scale into this regime. We present results for two popular discretizations of the Dirac operator, Wilson-clover and improved staggered, employing up to 256 GPUs on the Edge cluster at Lawrence Livermore National Laboratory.
Open AccessScaling Lattice QCD beyond 100 GPUs
Over the past five years, graphics processing units (GPUs) have had a transformational effect on numerical lattice quantum chromodynamics (LQCD) calculations in nuclear and particle physics. While GPUs have been applied with great success to the post-Monte Carlo "analysis" phase which accounts for a substantial fraction of the workload in a typical LQCD calculation, the initial Monte Carlo "gauge field generation" phase requires capability-level supercomputing, corresponding to O(100) GPUs or more. Such strong scaling has not been previously achieved. In this contribution, we demonstrate that using a multi-dimensional parallelization strategy and a domain-decomposed preconditioner allows us to scale into this regime. We present results for two popular discretizations of the Dirac operator, Wilson-clover and improved staggered, employing up to 256 GPUs on the Edge cluster at Lawrence Livermore National Laboratory.
- Open Access
Parallelizing the QUDA Library for Multi-GPU Calculations in Lattice Quantum Chromodynamics
Graphics Processing Units (GPUs) are having a transformational effect on numerical lattice quantum chromodynamics (LQCD) calculations of importance in nuclear and particle physics. The QUDA library provides a package of mixed precision sparse matrix linear solvers for LQCD applications, supporting single GPUs based on NVIDIA's Compute Unified Device Architecture (CUDA). This library, interfaced to the QDP++/Chroma framework for LQCD calculations, is currently in production use on the "9g" cluster at the Jefferson Laboratory, enabling unprecedented price/performance for a range of problems in LQCD. Nevertheless, memory constraints on current GPU devices limit the problem sizes that can be tackled. In this contribution we describe the parallelization of the QUDA library onto multiple GPUs using MPI, including strategies for the overlapping of communication and computation. We report on both weak and strong scaling for up to 32 GPUs interconnected by InfiniBand, on which we sustain in excess of 4 Tflops.
Open AccessParallelizing the QUDA Library for Multi-GPU Calculations in Lattice Quantum Chromodynamics
Graphics Processing Units (GPUs) are having a transformational effect on numerical lattice quantum chromodynamics (LQCD) calculations of importance in nuclear and particle physics. The QUDA library provides a package of mixed precision sparse matrix linear solvers for LQCD applications, supporting single GPUs based on NVIDIA's Compute Unified Device Architecture (CUDA). This library, interfaced to the QDP++/Chroma framework for LQCD calculations, is currently in production use on the "9g" cluster at the Jefferson Laboratory, enabling unprecedented price/performance for a range of problems in LQCD. Nevertheless, memory constraints on current GPU devices limit the problem sizes that can be tackled. In this contribution we describe the parallelization of the QUDA library onto multiple GPUs using MPI, including strategies for the overlapping of communication and computation. We report on both weak and strong scaling for up to 32 GPUs interconnected by InfiniBand, on which we sustain in excess of 4 Tflops.
- Open Access
The role of multigrid algorithms for LQCD
We report on the first successful QCD multigrid algorithm which demonstrates constant convergence rates independent of quark mass and lattice volume for the Wilson Dirac operator. The new ingredient is the adaptive method for constructing the near null space on which the coarse grid multigrid Dirac operator acts. In addition we speculate on future prospects for extending this algorithm to the Domain Wall and Staggered discretizations, its exceptional suitability for high performance GPU code and its potential impact on simulations at the physical pion mass.
Open AccessThe role of multigrid algorithms for LQCD
We report on the first successful QCD multigrid algorithm which demonstrates constant convergence rates independent of quark mass and lattice volume for the Wilson Dirac operator. The new ingredient is the adaptive method for constructing the near null space on which the coarse grid multigrid Dirac operator acts. In addition we speculate on future prospects for extending this algorithm to the Domain Wall and Staggered discretizations, its exceptional suitability for high performance GPU code and its potential impact on simulations at the physical pion mass.
- Open Access
Moebius Algorithm for Domain Wall and GapDW Fermions
The M\"obius domain wall action \cite{Brower:2004xi} is a generalization of Shamir's action, which gives exactly the same overlap fermion lattice action as the separation ($L_s$) between the domain walls is taken to infinity. The performance advantages of the algorithm are presented for small ensembles of quenched, full QCD domain wall and Gap domain wall lattices \cite{Vranas:2006zk}. In particular, it is shown that at the larger lattice spacings relevant to current dynamical simulations M\"obius fermions work well together with GapDWF, reducing $L_s$ by more than a factor of two. It is noted that there is a precise map between the domain wall and effective overlap action at finite quark mass including finite $L_s$ chiral violations so that the Ward-Takahashi identities for the axial and vector currents are exactly equivalent in the two formulations.Open AccessMoebius Algorithm for Domain Wall and GapDW Fermions
The M\"obius domain wall action \cite{Brower:2004xi} is a generalization of Shamir's action, which gives exactly the same overlap fermion lattice action as the separation ($L_s$) between the domain walls is taken to infinity. The performance advantages of the algorithm are presented for small ensembles of quenched, full QCD domain wall and Gap domain wall lattices \cite{Vranas:2006zk}. In particular, it is shown that at the larger lattice spacings relevant to current dynamical simulations M\"obius fermions work well together with GapDWF, reducing $L_s$ by more than a factor of two. It is noted that there is a precise map between the domain wall and effective overlap action at finite quark mass including finite $L_s$ chiral violations so that the Ward-Takahashi identities for the axial and vector currents are exactly equivalent in the two formulations. - Open Access
Strange quark content of the nucleon
We discuss the calculation of disconnected diagrams needed for determining the strange quark content of the nucleon on the lattice. We present results for the strange scalar form factor and the related parameter f_Ts, which enters into the cross-section for the scattering of dark matter off nuclei in supersymmetric extensions of the standard model. In addition, we present results for the strange contribution to the nucleon's axial and electromagnetic form factors. The calculations were performed with two dynamical flavors of Wilson fermions on a 24^3 x 64 anisotropic lattice with a_s = 3a_t = 0.11 fm and M_pi ~ 400 MeV.
Open AccessStrange quark content of the nucleon
We discuss the calculation of disconnected diagrams needed for determining the strange quark content of the nucleon on the lattice. We present results for the strange scalar form factor and the related parameter f_Ts, which enters into the cross-section for the scattering of dark matter off nuclei in supersymmetric extensions of the standard model. In addition, we present results for the strange contribution to the nucleon's axial and electromagnetic form factors. The calculations were performed with two dynamical flavors of Wilson fermions on a 24^3 x 64 anisotropic lattice with a_s = 3a_t = 0.11 fm and M_pi ~ 400 MeV.
- Open Access
Blasting through lattice calculations using CUDA
Modern graphics hardware is designed for highly parallel numerical tasks and provides significant cost and performance benefits. Graphics hardware vendors are now making available development tools to support general purpose high performance computing. Nvidia's CUDA platform, in particular, offers direct access to graphics hardware through a programming language similar to C. Using the CUDA platform we have implemented a Wilson-Dirac operator which runs at an effective 68 Gflops on the Tesla C870. The recently released GeForce GTX 280 runs this same code at 92 Gflops, and we expect further improvement pending code optimization.
Open AccessBlasting through lattice calculations using CUDA
Modern graphics hardware is designed for highly parallel numerical tasks and provides significant cost and performance benefits. Graphics hardware vendors are now making available development tools to support general purpose high performance computing. Nvidia's CUDA platform, in particular, offers direct access to graphics hardware through a programming language similar to C. Using the CUDA platform we have implemented a Wilson-Dirac operator which runs at an effective 68 Gflops on the Tesla C870. The recently released GeForce GTX 280 runs this same code at 92 Gflops, and we expect further improvement pending code optimization.
- Open Access
Strange quark contribution to nucleon form factors
We discuss methods for the calculation of disconnected diagrams and their application to various form factors of the nucleon. In particular, we present preliminary results for the strange contribution to the scalar and axial form factors, calculated with N_f=2 dynamical flavors of Wilson fermions on an anisotropic lattice.
Open AccessStrange quark contribution to nucleon form factors
We discuss methods for the calculation of disconnected diagrams and their application to various form factors of the nucleon. In particular, we present preliminary results for the strange contribution to the scalar and axial form factors, calculated with N_f=2 dynamical flavors of Wilson fermions on an anisotropic lattice.