Hello, A yearly allocation for the LCRC cluster has been requested with the following updated information: Submitter/PI: Robert Wiringa Project Name: qmc_for_nuclei Division: PHY Project title: Quantum Monte Carlo Calculations of Light Nuclei Associated funding: DOE Office of Nuclear Physics Other Systems: ANL Blue Gene/Q (INCITE 57M hours for 12C studies) ANL Theta (ALCC 35M hours for chiral nuclear spectra) Science: This project uses Quantum Monte Carlo [Green's function (GFMC), variational (VMC), cluster VMC, and auxiliary-field diffusion (AFDMC)] methods to compute ground-state and low-lying excited-state expectation values of energies, densities, electroweak transitions and response, spectroscopic overlaps, etc., for light- and medium-mass nuclei and low-energy scattering reactions involving these nuclei. Realistic two- and three-nucleon potentials and one- and two-body electroweak charge and current operators are used. Our goal is a description of all of these systems using a Hamiltonian that also provides an excellent description of nucleon-nucleon scattering and nucleonic matter. Such a `standard nuclear model' can then be used, for example, to compute low-energy astrophysical reactions which cannot be experimentally measured. We are also actively engaged in making predictions that can and are being tested at experimental facilities such as ATLAS, NSCL, TRIUMF, and JL ab. Project description: A) Nuclear Structure with chiral effective field theory We have developed new nuclear Hamiltonians, including two- and three-nucleon (NN+3N) potentials, based on chiral effective field theory (XEFT) including Delta-isobar degrees of freedom and formulated in configuration space, in the last few years. XEFT potentials have become a popular way to formulate nuclear forces, but up to now they were only in momentum space, and not amenable to our very accurate quantum Monte Carlo methods which require predominantly local interactions. We have recently calculated in GFMC (VMC) more than 35 (100) nuclear ground and excited states up to A=12 for one of the new models, using both Blues and Theta. We have three more models to evaluate, and these will require a search to fit the 3N potential parameters to get the best reproduction of nuclear spectra. We also will be investigating higher-order 3N potentials. Larger nuclei will be computed on Theta, but Bebop will be more efficient for the lighter (A<10) nuclei. We request 300,000 cpu ho urs in FY18 for this purpose. B) A=11,12 nuclei Extensive effort has gone into GFMC calculations of 0+ states in 12C using Mira, but only a few exploratory calculations of other A>10 nuclei have been made. We have recently made progress on trial functions for 11B and 12Be, particularly with the help of optimization routines for the VMC input wave functions. However other cases like 11Be and 12B remain to be studied, particularly with wave functions that are "clusterized." The VMC codes for these trial functions are running well using a mix of OpenMP inside nodes and MPI between nodes, but some steps in the calculation can only use about 12-16 MPI ranks (with 8-16 OMP threads) efficiently. These steps are best performed on Bebop (rather than Theta). We will be using both our standard AV18+IL7 Hamiltonian and the newer chiral Hamiltonians mentioned above. We would like 300,000 cpu hours for this work in FY18. C) Weak decays of light nuclei We are studying the weak decays of light nuclei, including 3H, 6He, (7,8,9)Li, 7Be, 8B, and (9,10,11)C. A paper on 'easy' allowed and superallowed Fermi and Gamow-Teller transitions is being completed, but more difficult transitions remain to be evaluated. The calculations include both one- and two-body current operators consistent with our standard AV18+IL7 Hamiltonian. However we also need to investigate the new XEFT potentials. This work is relevant to ongoing experiments at ATLAS and at U. of Washington. We are also making model studies of neutrinoless double beta decay, a top priority as the next major DOE Nuclear Physics construction project, in (8,10)He and 12Be. We would like 200,000 cpu hours for this work in FY18. D) Neutron matter The auxiliary field diffusion Monte Carlo (AFDMC) method is a more efficient algorithm for dealing with larger nuclear systems compared to the GFMC. We will calculate the equation of state (EoS) of neutron matter by using AFDMC for the new XEFT interactions. The EoS of neutron matter is the key ingredient for predicting the structure and other properties of neutron stars. By performing calculations with different cutoffs regulating the short-range part of the interactions, we will give an estimate of theoretical uncertainties in the EoS as a function of the density. We request 200,000 cpu hours in FY18 for this work. Industry partnership: The URL window might not accept our project website, so I am listing it here: www.phy.anl.gov/theory www.phy.anl.gov/theory/research/av18/ www.phy.anl.gov/theory/research/QMCresults.html Project URL: http://www.phy.anl.gov/theory/research/QMCresults.html Current FY Hours Used: undetermined amount New FY Requested allocation: 1000000 Q1: 250000 Q2: 250000 Q3: 250000 Q4: 250000 Justification: Our QMC programs scale very well both with increasing number of OpenMP threads and MPI ranks. For example a GFMC job run on Bebop used 100 Broadwell nodes, each with 2 MPI ranks. Each rank used 16 OpenMP threads and 1 Pthead for memory management. The job ran for 110 minutes. The MPI efficiency (which we always measure by separately timing computational loops and communication calls) was 93.6%. This means that, on average, each rank spent 93.6% of its wall time doing calculations. An OpenMP scaling study of the wave function subroutines on a Broadwell node using avx2 used 1635, 908, 501, and 342 milliseconds for 2, 4, 9, and 18 threads. A KNL node used 1979 and 1210 milliseconds for 8 and 16 threads. Based on these results the above large run would now be made using 4 ranks/node with 8 threads each. We do not benefit much from the vector instructions. Using avx512 on the KNL nodes increased the subroutine speed by less than 2% compared to avx2. We find the KNL nodes to be much less efficient than the Broadwell ones. For example, using 4 ranks/node with 9 threads/rank on Broadwell or 16 threads/rank on KNL gives 40% less throughput on the KNL nodes. Storage requirements: no additional space required Thank You, The LCRC Accounts System