Keyword Suggestions

NN10.09 - Parallelization of Molecular-Dynamics Simulations Using Tasks 
December 4, 2014   4:30pm - 4:45pm

Molecular-dynamics (MD) simulation is one of the most important computational methods to study the properties of materials on the atomistic scale. While parallel MD codes have been available for a long time, computing technology has evolved. For the last years, the computational power of CPUs has been increased primarily through increases of the number of cores and SIMD units with wider vector registers. This presentation addresses the question of how MD simulation programs can take full advange of the computational power provided by modern CPUs.

A novel parallel algorithm for MD simulations with short-ranged forces is discussed. The cell-task method uses a task-based programming approach. The calculation of forces is divided into a large number of small tasks that are executed by a thread pool. This approach avoids the load balancing problems that can occur with the spatial decomposition method in simulations of inhomogeneous systems. Benchmark calculations are presented which show that the cell task method significantly outperforms spatial decomposition in the case of inhomogeneous, porous systems. For homogeneoeus systems both methods perform equally well. Tests on an Intel Xeon-Phi co-processor show that the method scales very well to large numbers of threads and is thus suitable for many-core processors with hundreds of CPU cores.

Recent CPUs contain SIMD vector units that allow simultaneous operations on short vectors (typically 2 - 8 elements). Unfortunately, the sparse and irregular interaction pattern in MD simulations with short-ranged forces makes it difficult to take advantage of SIMD vectorization. In the second part of this presentation a blocking algorithm is discussed which allows vectorization of some parts of the force calculation of short-ranged MD simulations. Bechmakr caclulations showing signifcant speed-up factors of two and above will be presented.

Average Rating: (No Ratings)
  Was great, surpassed expectations, and I would recommend this
  Was good, met expectations, and I would recommend this
  Was okay, met most expectations
  Was okay but did not meet expectations
  Was bad and I would not recommend this

Performance Enhancement of Pentacene Based Organic Field-Effect Transistor through DNA Interlayer
Semiconducting Polymer-Dipeptide Nanostructures by Ultrasonically-Assisted Self-Assembling
DNA as a Molecular Wire: Distance and Sequence Dependence
Structure-Property Relationship in Biologically-Derived Eumelanin Cathodes Electrochemical Energy Storage
Artificial Physical and Chemical Awareness (proprioception) from Polymeric Motors