PRACE Preparatory Access - cut-off evaluation in September 2011

Find below the results of the cut-off evaluation of the 1st of September 2011 for the PRACE Preparatory Access Calls.

Type A - Code scalability testing

 Project name: Performance and Scalability of HadGEM2-ES and HadGEM3 Configurations of the UK Met. Office Unified Model

Project leader: Grenville Lister, National Centre for Atmospheric Science, Reading, UK
Collaborators: Simon Wilson, National Centre for Atmospheric Science, Reading, UK
Research field: Earth Sciences and Environment

Abstract Much of the UK atmospheric science and Earth system modeling community use as one of their principal codes, the Unified Model (UM) developed by the UK Met. Office in collaboration with partners in academe. Several versions of the code, each tuned to best represent specific areas of research interest are currently running on several different machine architectures. We propose to investigate the performance on CURIE of two UM configurations which we believe will be of particular importance in the future of Earth system modeling and climate science.

The first configuration will be HadGEM2 (Hadley Centre Global Environmental Model version 2), the Earth system model including fully interactive ocean and chemistry components which is currently used by several centres worldwide to provide data for CMIP5 (5th Climate Model Inter-Comparison Project) and is the basis for several current research studies and will be the basis for an increasing number in the future as modelers wish to benefit from both the increased fidelity and resolution that HadGEM2 provides over earlier configurations. The second will be HadGEM3 which is currently under development and represents the state-of-the-art both in terms of its representation of atmospheric processes and its unprecedented high resolution. HadGEM2 currently runs on hundreds of processors while HadGEM3 currently runs on more than 6000 processors (with MPI and OpenMP) on a Cray xe6. Both models will be ported to CURIE along with ancillary data and will be tested to determine their scaling and performance characteristics in respect of running on differing numbers of processors, different processor configurations, and on both fat and thin node partitions.

Since HadGEM2, with its very high level of complexity, will be the production Earth system model of the future for many in the UK academic community, it is essential that we have a clear understanding of its architectural sensitivities and can advise the modeling community of its performance on the Bull machine. The case for understanding and documenting the behaviour of the very high resolution HadGEM3 on CURIE is as compelling given its significantly increased computational resource requirements.

Computer system: CURIE, GENCI/CEA
Resource awarded: 50 000 core-hours

---

 Project name: Analysis of Large Conformational Transitions in Protein Kinases by Enhanced-Sampling Molecular Dynamics Simulations

Project leader: Amedeo Caflisch, University of Zurich, Zurich, Switzerland
Collaborators: Ting Zhou, University of Zurich, Zurich, Switzerland
Research field: Medicine and Life Sciences

Abstract For efficient sampling by atomistic MD simulations, we have developed a novel method based on complex network analysis and free-energy profiles. Our method speeds up the exploration of conformational space by 1 to 2 orders of magnitude, preserves the free energy barriers, and does not need reweighting. Due to its intrinsic parallel character, it is well suited for running on modern computer clusters. It has been successfully applied to a wide range of simulation systems from the folding of a beta-sheet mini protein (whose configuration space network consists of more than half a million microstates) to coarse-grained simulations of fibril formation. Our approach can be straightforwardly transferred to explicit-solvent MD simulations of large (multi-domain) proteins. We propose to characterize the large-scale conformational transitions in protein kinases in atomistic detail. In this way, highly-populated metastable states on the free energy surface will be identified and used for discovering novel types of kinase inhibitors. Protein kinases modulate a spectrum of biological processes by their phosphorylation activities. All kinases have a conserved activation loop, which can adopt a large number of conformations. Atomistic studies of kinases have traditionally been carried out in the active state, in which the (phosphorylated) activation loop adopts the enzymatically competent conformation. In the inactive state, the activation loop, however, blocks the substrate binding site. Due to the movement of the so-called DFG motif in the activation loop, and in particular the Phe side chain, a hydrophobic pocket becomes empty. This allosteric site provides additional handle for tuning potency and selectivity. Allosteric inhibitors have been sought for only a small fraction of the kinome. The inactive state of protein kinases is usually detected when stabilized by a ligand that addresses the allosteric pocket or in the auto inhibited state. This makes extremely difficult targeting the inactive kinase with undisclosed allosteric ligands and an unidentified auto inhibited state. The development of selective, allosteric inhibitors of protein kinases relies on a thorough understanding of the dynamics in the cleft region that includes not only the activation loop but also the glycine-rich loop. The quantitative analysis of the free energy landscape is not easily accomplished in experiments or by computational methods. From the computational point of view, the main difficulty arises from slow processes, i.e., conformational transitions involving high free energy barriers, and insufficient parallel efficiency of molecular dynamics simulations (including brute force) for attaining complete and accurate results within an affordable time. Our enhanced sampling approach is ideally suited for analyzing slow transitions in (biological) macromolecules and will provide us with a clear picture of the pharmacologically relevant conformations in protein kinases.

Computer system: CURIE, GENCI/CEA
Resource awarded: 50 000 core-hours

Computer system: HERMIT Gauss/HLRS:
Resource awarded: 50 000 core-hours

---

 Project name: Large-scale O(N) DFT simulations of defects in metal oxides

Project leader: Pérez Rubén, Universidad Autónoma de Madrid, Madrid, Spain
Collaborators: Milica Todorovic, Universidad Autónoma de Madrid, Madrid, Spain
Research field: Chemistry and Materials

Abstract Investigation of novel catalytically active surfaces requires a comprehensive experimental method for the identification and rapid characterisation of prospective catalytically active sites and local probes have proved themselves particularly suitable to this task. The powerful new method of three-dimensional atomic force microscopy (3D-AFM) in non-contact mode has recently been combined with scanning tunneling microscopy (STM) to study the oxygen-terminated copper (100) surface. Complex 3D data sets, obtained by simultaneously recording the tunneling current and the AFM frequency shift, allow for site-specific quantification of forces and tunneling currents. The wealth of information obtained is promising for future applications, but the interpretation of the wide range of contrast modes requires a thorough characterisation of the sources of contrast in AFM and STM imaging.

We combine density-functional theory (DFT) total-energy calculations with Non-equilibrium Green’s Function (NEGF) methods for electronic transport to clarify the different contrast modes obtained in the experiments, explore surface properties, and advance the understanding and control of this new experimental method. AFM simulations performed by applying tips of different reactivity to the Cu-O surface explore the site-specific tip-surface interactions and any associated structural deformations.. Subsequent charge density and STM simulations reveal the range of contrasts associated with the varying electronic properties of the tips.

The combined method can be applied to clean surfaces, as well as surface steps, kinks and defects to aid experimental recognition of surface features. To this purpose, our calculations span different length-scales to describe very local surface effects, as well as long-range defects or surface domain structures. We employ a linear-scaling DFT method implemented in the OpenMX code to simulate large surface reconstructions with a high degree of accuracy. Such calculations are particularly challenging as there are very few successful applications of this theoretical method to metallic systems, but the nature of the Krylov subspace method implementation of O(N) DFT is uniquely suitable to our needs. Spanning different length-scales, the combination of conductance calculations with total energy methods provides insight into (1) the fundamentals of contrast formation in this novel experimental technique and (2) into the correlation between tip sample forces and local chemical reactivity, factors that are essential for the further development and application of this novel approach to characterise catalytic activity.

This work is carried out in collaboration with experimental groups of Prof. U. D. Schwarz and Prof. E. I. Altman at Yale University, USA and with Prof. T. Ozaki at JAIST, Japan. It is supported by a NSF-MICINN grant.

Computer system: CURIE, GENCI/CEA
Resource awarded: 50 000 core-hours

---

 Project name: Highly Scalable Simulations of Complex, Multicomponent Biomembrane Systems

Project leader: Mark Sansom, University of Oxford, Oxford, UK
Collaborators: Antreas Kalli / Joseph Goose, University of Oxford, Oxford, UK
Research field: Medicine and Life Sciences

Abstract Membrane proteins account for the 25% of all genes and they are involved in various diseases ranking from diabetes to cancer. They also play an important role in many cellular processes such as signal transduction, transport and cell-cell interactions. Recently, the number of known structures of membrane proteins is increased, however the conformational dynamics of the static structure needs to be studied further to better understand their biological function. Molecular dynamics simulations allow the study of the dynamics of proteins however, especially simulations of membrane proteins, usually contain millions of atoms and therefore the use of supercomputers, like PRACE resources, are vital in studying this complexes.

The overall aim of this project is to stretch the envelope with respect to biomolecular simulations as applied to systems of biomedical importance. This will be achieved by multi-scale simulation of large scale biomembrane systems. We have experience of running large (up to 4.5million particle) vesicle and bilayer molecular dynamics simulations, on a Cray XT4 (JADE) computer showing reasonable scaling up to 1024 cores and more recently up to 3072 cores on a Cray XE6 system (HECToR – phase 2b). These simulations were using a coarse-grained protocol. Coarse-grained MD (CG-MD) simulations represent groups of atoms as effective particles, resulting in speed-ups of about two to three orders of magnitude compared to all-atom simulations. CG simulations allow the study of assembly and aggregation of macromolecular systems. Recent endeavours within our group have led to the need to simulate larger multi-body atomic systems in addition to coarse-grained for which different run parameters and are necessary. We have a series of systems to optimise and benchmark covering a wide range of biomembrane system which we hope to use in support of a future PRACE Tier-0 access application.

Computer system: CURIE, GENCI/CEA
Resource awarded: 50 000 core-hours

Computer system: Jugene, Gauss/FZJ
Resource awarded: 100 000 core-hours

Computer system: HERMIT Gauss/HLRS:
Resource awarded: 50 000 core-hours

---

 Project name: Large scale MD simulations of nanosomes under electric field

Project leader: Mounir Tarek, CNRS-Nancy University, Vandoeuvre-lès-Nancy, France
Collaborators: Lucie Delemotte, Nancy University, Vandoeuvre-lès-Nancy, France
Research field: Chemistry and Materials

Abstract Currently, computational approaches remain potentially the only techniques able to follow, at the atomic scale the local perturbation lipid membranes undergo when they are subject to external electric field. In this proposal, we aim to port and investigate scalability and performance of 2 open source codes, NAMD a US based code and GROMACS a Europeen code, able today to handle systems of over a Million particle. We have last year used NAMD on JADE a French Tier 1 machine to study effects of electric fields on lipid liposomes (between 1.2 and 3 Million Atoms) with limited success (efficient scaling not exceeding 2024 cores). We plan here to assess and compare for the first time (to the best of our knowledge) the efficiency of these codes for systems of such size. The main aim here is to obtain scalability plots which can be used as supporting information when applying to future PRACE project.

Computer system: CURIE, GENCI/CEA
Resource awarded: 50 000 core-hours

Computer system: HERMIT Gauss/HLRS:
Resource awarded: 50 000 core-hours

---

 Project name: Developing improved models for Tsunamis

Project leader: Antonio Puertas Gallardo, European Commission, Joint Research Centre, Ispra, Italy
Collaborators: Alessandro Annunziato / Giovanni Franchello, European Commission, Joint Research Centre, Ispra, Italy
Research field: Earth Sciences and Environment

Abstract The objective of the project is to optimize a calculation procedure that allows to perform in the fastest possible time, calculations related to the development of Tsunami Early Warning Systems.

The calculations will be performed with coupled codes (SWAN and HYFLUX) which are needed to evaluate under several potential initial conditions, the estimate of the damage on the coasts as a consequence of Tsunami events.

Computer system: CURIE, GENCI/CEA
Resource awarded: 50 000 core-hours

Type B – Code development and optimization by the applicant (without PRACE support)

---

 Project name: HPC Methods and Applications for Highly Scalable Molecular Dynamics Simulation

Project leader: Jadran Vrabec, University of Paderborrn, Paderborn, Germany
Collaborators: Martin Horsch, Technical University of Kaiserslautern, Kaiserslautern, Germany; Wolfgang Eckhardt, Technical University of München, Garching, Germany; Stefan Eckelsbach, University of Paderborrn, Paderborn, Germany
Research field: Chemistry and Materials

Abstract Molecular simulation allows the study of phenomena, where classical approaches like continuum methods fail. The range of application fields is very broad and reaches from materials science, over process engineering to bio- and nano-engineering. Numerous practical problems can be tackled by molecular dynamics (MD) and Monte Carlo (MC) simulations on the basis of molecular force fields that were hardly accessible before. The importance of the molecular simulation methods for the future scientific and industrial research and development has widely been understood, however, applications in industry are still rare. The main reason is the large numerical effort that is associated with these methods. This hinders on the one hand the development of realistic force fields and, on the other hand, their application in industrial organizations. The advent of massively parallel molecular simulation programs and the according hardware will lead to a breakthrough in the near future.

This project aims at the testing of the scalability of a massively parallel molecular simulation code named ls1-Mardyn. It has been developed in a joint effort by scientists from engineering and informatics within the last few years and was recently published in scientific journals (Journal of Computational Science, doi:10.1016/j.jocs.2011.01.009).

ls1-Mardyn aims at large scale molecular dynamics studies of nano-scale phenomena in fluids. The scalability of ls1-Mardyn has been shown for up to 2048 processes on a NEC Nehalem cluster. In the present project, it is intended to test and identify performance bottlenecks and to optimize the scalability on other massively parallel hardware platforms, to allow for efficient program execution on petaflop architectures.

Computer system: CURIE, GENCI/CEA
Resource awarded: 200 000 core-hours

Computer system: JUGENE, Gauss/FZJ
Resource awarded: 250 000 core-hours

Type C – Code development with support from experts from PRACE

 Project name: Design and implementation of an efficient Parallel Monte Carlo (MC) algorithm for the atomistic simulation of nanostructured materials

Project leader: Vlasis Mavrantzas, University of Patras, Patras, Greece
Collaborators: Efthymios Housos / Vasileios Kolonias / Kyriakos Stefanidis / Dimitris Tsalikis / Orestis Alexiadis, University of Patras, Patras, Greece
Research field: Chemistry and Materials

Abstract Materials which exhibit ordered morphology at the nano-scale have drawn considerable attention in the last two decades due to their unique combination of opto-electronic properties, ease of preparation and low cost manufacturing. Systems like semiconducting polymers (e.g., semi-crystalline poly-thiophenes), micro-crystalline silica and graphene have proved to be very promising candidates for a variety of applications like organic micro-electronics and photovoltaics. Simulating these materials at the nanoscale is inefficient or even unfeasible with Molecular Dynamic (MD) methods because of the problem of long relaxation times, especially when highly ordered structures form at low enough temperatures. In order to circumvent the MD drawbacks, it’s imperative to develop and implement efficient Monte Carlo (MC) techniques (Metropolis Monte Carlo and Kinetic Monte Carlo) to effectively simulate systems with large scale nanophase-separated structures and overcome obstacles related with large system sizes and sluggish dynamics. Conventional Monte Carlo schemes, however, typically encompass moves which become less and less effective when the simulation is conducted at low temperatures, in the sense that their acceptance rate becomes smaller and smaller. This presents a considerable problem when one wishes to simulate the system over a wide range of temperatures and especially at the lower ones where the crystalline nano-domains are formed, since for the simulation to be ergodic, the MC moves should be characterized by sufficiently high acceptance rates. To deal with such a problem, we resort to parallel versions of the Monte Carlo technique, built on the parallel tempering method1 which can be realized both for different temperatures (parallel tempering) (Ti ,…, Tn with T1 > T2 > … > Tn) but also for different values of the parameter σ describing repulsive core potential interactions (parallel excluded volume2) (σi ,…, σn with σ1 > σ2 > … > σn) at the same temperature. We propose here to design two different editions of such a parallel MC algorithm in order to simulate two very important families of systems today: polymer semiconductors based on alkylthiophenes such as poly(3-Hexylthiophene) (P3HT) and micro-crystalline silicon thin films. The interest in these materials stems from the demand nowadays for low-cost, large-area semiconducting devices for displays and photovoltaic applications which is also reflected in the value of the market for devices for use in microelectronics; according to the European Semiconductor Industry Association, http://www.eeca.eu/index.php/esia_h..., this exceeds 12 billion € today.

Computer system: CURIE, GENCI/CEA
Resource awarded: 200 000 core-hours

---

 Project name: Atmospheric research and operational weather prediction

Project leader: Alexei Strelchenko, The Cyprus Institute, Nicosia, Cyprus
Collaborators: Mohamed Abd El-Kader / Patrick Fitzhenry, The Cyprus Institute, Nicosia, Cyprus
Research field: Earth Sciences and Environment

Abstract The Weather Research and Forecasting (WRF) model is a mesoscale numerical weather prediction and atmospheric simulation system designed for both research and operational applications. It features multiple dynamical cores, a 3-dimensional variational data assimilation system (3DVAR) , and a software architecture allowing for computational parallelism and system extensibility.

The WRF Software Framework provides the infrastructure that accommodates two dynamics solvers, the Advanced Research WRF (ARW) and Non-hydrostatic Mesoscale Model (NMM), multiple physics packages that plug into the solvers through a standard physics interface, variational data assimilation system (WRF-Var) and air chemistry modeling system (WRF-Chem).

The WRF physics options fall into several categories that include micro-physics, cumulus parametrization, planetary boundary layer, land-surface model and radiation.

WRF is suitable for a broad spectrum of applications across scales ranging from meters to thousands of kilometers. Such applications include real-time numerical weather prediction, data assimilation development and studies, parametrized-physics research, regional climate simulations, air quality modeling, atmosphere-ocean coupling and idealized simulations (e.g., convection, baroclinic waves etc.).

Computer system: CURIE, GENCI/CEA
Resource awarded: 200 000 core-hours

Computer system: HERMIT Gauss/HLRS:
Resource awarded: 200 000 core-hours