June 1st, 2011UncategorizedJuan Sebastian Totero Gongora 0 Comments

Recently, contributing to the 2009 annual survey of the EDGE foundation entitled “What will change everything”, T. Sejnowski foresees that the computers will be the “microscopes of the future” and, specifically, that are computers that “have made it possible to localize single molecules with nanometer precision and image the extraordinary complex molecular organization inside cells”. In his contribution Sejnowski was recognizing the importance of automatic and computerized control of laser beam and image analysis in modern optical microscopy. However, more importantly, computers can be also considered as “future microscopes” for the capability of simulating at the atomic scale the behavior of matter and biological systems. To reach this goal, which was even unthinkable up to some decades ago, it is necessary to develop software able to simulate the newtonian dynamics of a large number of atoms (order 10^{12}, the “number” of atoms in a cell) for long enough time (order microsecond, or 10^{10} time steps). The state-of-the art supercomputers, on the hardware side, and the parallel simulation codes, on the software one, are nowadays on the way to get this result. However, before facing the ultimate problem of simulating the complexity of life (micro-)organisms, one should validate and optimize codes that simulate hard (physical and chemical) systems. Even if smaller than a micro-organism, these systems encompass problems which are extremely demanding in terms of computational resources, with e.g., simulation boxes containing pico-molar quantity of matter and with characteristic times of 0.1 microseconds. Molecular Dynamics (MD) is expected to be the key to efficiently solve these type of problems.


With reference to molecular dynamics, several methods have been published in the past which incorporate different degrees of parallelism. To date, MD scaling has been demonstrated up to ten thousand of cores, with a speed of about 7 iterations per second for an ensemble of 1 billion particles running on 65536 cores of a BlueGene/L. For several applications of molecular dynamics, such as the study of structural glasses where a typical simulation requires 10^{6-7}  time steps, this speedup is still too small to perform practical calculations. In the study of glasses, and in the more general area of amorphous materials, molecular dynamics simulations are extremely important as they are able to glimpse the system dynamics at spatial scales between 1-100 nm, a range that is completely inaccessible with experimental apparatus. However, due to such speed bottleneck problems, present MD studies on glass have been limited to a number of particles (10 million) unable to have simulation boxes large enough to capture the interesting phenomenology at the micron-scale.


For this reason, I have recently developed a Billions Body Molecular Dynamics (BBMD) package, and demonstrate its effectiveness in the study (as case study) of structural glasses by analyzing the glass formation of an exceptionally-large particles system. The BBMD code was able to scale on the whole  294912 cores of the BlueGene/P system at the Julich Supercomputing Centre, which constitutes the world’s largest supercomputer available characterized by 72 racks of an IBM BlueGene/P. In such an extreme scaling test, the BBMD code showed an efficiency of about 90%, and a overall speed of 2 seconds x iteration with 100 billion particles.


Read the full Article on http://dx.doi.org/10.1016/j.jcp.2012.01.019