Parallel programming in script languages can be educational and fun and while there are multiple MPI wrappers for Python (at different levels of sophistication and completeness), an equivalent Tcl wrapper seems to be missing (except for the very minimal proof of concept here). With Tcl being the most pervasively used scripting language in VMD and data sets getting larger while analysis scripts become more sophisticated, there is a potential market for it. In addition, writing wrappers for external libraries in Tcl is fairly straightforward. So here we have TclMPI, the MPI wrapper package that the world hasn't been waiting for. Uppdate: a new version has been uploaded which includes bugfixes, cleanups, and an integrated test suite.

The VMD/molfile plugin for HOOMD has been updated to the latest version and also all bundled binaries updated. The PowerPC MacOS X variant is not longer available since I have no longer access to such a machine. As a new addition, the bundled makefile has support for compiling a 32-bit windows plugin to using the MinGW cross-compiler suite that is available with Fedora Linux. Download the plugin package from here

Work has started to port the "colvars module" written by current ICMS member Giacomo Fiorin and former ICMS member Jerome Henin from NAMD to LAMMPS. Since the code was written with portability in mind, it turned out to be a minor burden and a first test release is available inside of LAMMPS-ICMS.

After halting the development of LAMMPS-ICMS to provide a stable version of LAMMPS with multi-threading support for several months, new updates have been pushed to the git repository again. The new USER-OMP package, which is a rewrite of the USER-OPENMP package, has reached the same level of maturity, is part of regular LAMMPS and offers overall much improved parallel efficiency and convenience. USER-OMP avoids many reduction operation and thus can also efficiently multi-thread bond and angle interactions. Also, the neighbor list generation has been adapted for multi-thread support and further parallel efficiency. LAMMPS-ICMS will continue to provide additional/experimental features and updates to USER-OMP and GPU and other packages before they get merged into LAMMPS. A few current examples: a simple loadbalancing scheme implemented by christoph kloss, a refactored and more detailed CPU and Wall time accounting and output, additional multi-threaded styles.

New snapshots for the HOOMD molfile plugin and topotools Tcl script plugin were uploaded. A new release (1.9.1) of VMD is planned for the near future, so please try and test them and report any bugs. Please note, that VMD 1.9 contains a bug that will prohibit reading of angle/dihedral/improper type labels even if they are properly listed in the xml file.

Thanks to the hard work of Mike Brown at ORNL and a hardware donation from AMD we have now a LAMMPS version with a GPU package that can be compiled for either CUDA or OpenCL and thus run on both Nvidia and AMD hardware. Without any specific tuning of the OpenCL kernels for the 4 ATI FirePro3D V8800 GPUs in our GPU test node (as shown in the attached pictures) the cards provide about 70-80% the acceleration that can be achieved with Nvidia Tesla C2050 GPUs in a similar setup.

 

Update: here is a first graph showing strong scaling for the LAMMPS "rhodo" benchmark for a CHARMM force field input of a rhodopsin molecule embedded in a lipid bilayer with surrounded by rigid TIP3P water. The y axis is the total time for 100 MD steps in seconds. LAMMPS has been compiled with single precision FFT-3.x support and mixed precision GPU support; both pair and kspace are GPU accelerated. Please note the double logarithmic scale. The same input takes on a single CPU core of the host machine (Xeon E5520, 2.27GHz 52.5 seconds, and with 8 MPI tasks 7.3 seconds. The ATI cards can beat this with a single GPU and two MPI tasks. The Nvidias currently still "win" with about 5x sustained speedup overall.

Over the last couple of weeks these two papers have been accepted for publication. High time to get busy on (finally) writing up all the OpenMP stuff in LAMMPS.

• Micellization studied by GPU-accelerated coarse-grained molecular dynamics, B.G. Levine , D.N. LeBard , R.H. DeVane , W. Shinoda , A. Kohlmeyer, M.L. Klein, J. Chem. Theory Comput. in press, (Link)
• Implementing Molecular Dynamics on Hybrid High Performance Computers - Particle-Particle Particle-Mesh, W.M. Brown, A. Kohlmeyer, S.J. Plimpton, A.N. Tharrington, Comp. Phys. Comm. in press  (Link)

While trying to add multi-threading to the LAMMPS classical MD code, i found a few changes that would make the corresponding serial code faster, too. In most cases the changes didn't make it into the regular pair styles (but of course will be in the USER-OMP package). In the case of the AIREBO pair style the improvements were quite significant. For a bulk CH test system (left side) with 5900 atoms the time for 100 time steps run with a serial executable dropped from 68 seconds to 31 seconds. For a test system with a carbon nanotube and four buckyballs with 940 atoms, the time for 1000 steps decreases from 21 seconds to 15 seconds. AIREBO is also one of the cases where parallelization via multi-threading is competitive with MPI parallelization. For the case on the left 8 MPI tasks require 4.8 seconds and 8 OpenMP threads 5.1 seconds; for the case on right 8 MPI tasks take much longer (5 seconds instead of 2.3 seconds) due to load imbalances caused by the domain decomposition, while the particle decomposition used in the OpenMP parallelization has no such issues. The OpenMP parallel AIREBO style will be released very soon, too. A very special thanks goes to Marcel Fallet and Steve Stuart at Clemson University for helping to test the correctness of the optimizations and providing additional bugfixes.