[gmx-developers] Re: GROMACS 4.6 pre-release: new algorithms, parallelization schemes, & GPU acceleration

[Szilárd Páll]

On Thu, Mar 1, 2012 at 12:59 AM, Szilárd Páll <szilard.pall at cbr.su.se> wrote:
> Dear Developers,
>
> With the GROMACS 4.6 release shaping up, I am writing to announce the
> public availability of some of the major features this release will
> introduce, and invite you for testing them:
> - "Verlet" cut-off scheme and the "nbnxn" non-bonded algorithms;
> - multi-level parallelization (MPI+OpenMP);
> - heterogeneous parallelization with native GPU-acceleration (MPI+OpenMP+CUDA).
>
> The code implementing the above features is brand new and requires
> intensive testing before reaching the beta state. The aim of this
> pre-beta phase is to uncover the remaining bugs and performance
> issues. To speed up the testing process, we would like to ask for your
> help with it.
>
> The code is in a stable, close to feature-complete state and resides
> in the nbnxn_hybrid_acc feature-branch during the pre-release testing.
> The pre-release phase effectively starts now and lasts until the code
> gets merged in to the 4.6 release branch. Fixes and improvements will
> be regularly pushed into the repository, so keep pulling for updates.
> Also, if you experience problems, the first thing you should do is to
> pull the changes and see if there are any new fixes that solve your
> issue.
>
>
> Before jumping at testing here's some documentation you should check out first:
> - Detailed description of the GROMACS cut-off schemes:
> http://www.gromacs.org/Documentation/Cut-off_schemes
> - Introduction to acceleration & parallelization in GROMACS:
> http://www.gromacs.org/Documentation/Acceleration_and_parallelization
> - CMake instructions:
> http://www.gromacs.org/Developer_Zone/Building_and_Testing/Cmake
> (Note that these wiki pages are still under construction; if you wish
> to help with improving them, contact us.)
>
> One-liner for the impatient ones:
> $ git clone git://git.gromacs.org/gromacs.git -b nbnxn_hybrid_acc &&
> cd gromacs && cmake ../ && make install -j4

Not that many have complained, but the one-liner is broken. A correct
alternative:
git clone git://git.gromacs.org/gromacs.git -b nbnxn_hybrid_acc && cd
gromacs && mkdir buid && cd build && cmake ../ && make install -j4

--
Szilárd

> Bug reports, suggestions on improvements, as well as bugfixes are
> welcome and much appreciated! Please consider submitting fixes to our
> gerrit code-review page (gerrit.gromacs.org), it's the quickest and
> easiest way contribute a patch for GROMACS!
>
> Cheers,
> Szilárd
>
> Footnote -- Known yet to be solved issues:
> - preliminary implementation of run configuration setup -- related
> mdrun options and environment variables will change;
> - preliminary (naive) NUMA awareness implementation -- can't use HT
> out of the box with (thread-)MPI+OpenMP;
> - manual CPU/core affinity setting needed for multiple process/machine
> runs (automatic thread pinning has to be turned off);
> - *avoid CUDA 4.1* for now; the GPU nbnxn kernels are, due to a
> regression in the nvcc compiler, 5-10% slower with CUDA 4.1 compared
> to 4.0;
> - CMake:
>  - no support for advanced CPU acceleration (AVX, SSE 4.1, native optimizations)
>  - no ICC 12 support
>  - not Windows support