Combining Task-based Parallelism and Adaptive Mesh Refinement Techniques in Molecular Dynamics Simulations

Abstract

Modern parallel architectures require applications to generate massive parallelism so as to feed their large number of cores and their wide vector units. We revisit the extensively studied classical Molecular Dynamics N-body problem in the light of these hardware constraints. We use Adaptive Mesh Refinement techniques to store particles in memory, and to optimize the force computation loop using multi-threading and vectorization-friendly data structures. Our design is guided by the need for load balancing and adaptivity raised by highly dynamic particle sets, as typically observed in simulations of strong shocks resulting in material micro-jetting. We analyze performance results on several simulation scenarios, over nodes equipped by Intel Xeon Phi Knights Landing (KNL) or Intel Xeon Skylake (SKL) processors. Performance obtained with our OpenMP implementation outperforms state-of-the-art implementations (LAMMPS) on both steady and micro-jetting particles simulations. In the latter case, our implementation is 4.7 times faster on KNL, and 2 times faster on SKL.

Adaptive grid implementation for parallel continuum mechanics methods in particle simulations

“…SFC-based domain decompositions yield an efficient partitioning scheme for stencil-like algorithms [200,201]. Adaptive variants, however, have not consistently proven to be beneficial [202][203][204], because of the more involved neighbor search. Additionally, in a short-range MD simulation, the main portion of the computational load is generated by the number of force pairs and only loosely coupled to the number of cells.…”

Optimization of the N-Body Simulation on Intel’s Architectures Based on AVX-512 Instruction Set

“…Regarding affinity, it can be seen that it is advisable to select one of the available strategies instead of delegating the distribution to the operating system (none). Unlike scatter, balanced and compact guarantee the proximity among OpenMP threads with consecutive identifiers, minimizing in this way the data communication that each thread requires 6 . As it was mentioned in Section 4.3, the compiler detects false dependencies in that loop and it is not able to generate SIMD binary code by itself.…”

“…Nowadays, the scientific community is experimenting with a new revolution on parallel processor technologies in the road to the Exascale. The novelties and enhancements not only involve hardware technologies but also changes in parallel programming models [6]. Beyond that, one of the most important challenges that still remains is how to perform large-scale simulations in a reasonable time using affordable computer systems.…”

“…The number of simulation steps was set to I = 100 5. A widely accepted convention in the available literature for this problem 6. Since all processor cores are in the same package, balanced and compact produce the same distribution when using all hardware threads…”

Towards Data-Flow Parallelization for Adaptive Mesh Refinement Applications

“…A communication thread per rank is used to coordinate the work-stealing, while OpenMP tasks conduct computations. Prat et al [22] studied the taskification of computations in AMR applications using OpenMP tasks and dependencies, combined with cache blocking and vectorization techniques. However, they did not include the study of communication patterns.…”