Transfer prallelisation information from old webpage to new one
Copy of Szilárds email here where we discussed what to do ...
As we discussed on Monday, the main documentation additions should go
into the user guide and website; not sure we need much if anything in
the manual given that nothing changes in terms of algorithms.
I think a review/rewrite of the old "Acceleration and parallelization"
page and the the partially overlapping "Getting good performance" page
of the user guide is a good start. This is where most of the PME-GPU
related content should go.
To be honest, I still prefer a lot of the original structure/content
that was left behind (or greatly shortened), but I am naturally
biased. Firstly, the intent of the original page was to explain the
different means of parallelization to the users (some of which is
referred to these days as acceleration). For this reason, the original
page first explained what "acceleration" refers to (i.e. accelerated
kernels like SIMD and GPU) followed by explanations of what different
means of parallelization are employed (MPI, OpenMP, MPI+OpenMP,
MPI+OpenMP+GPU, etc.), finally building up examples of how to
use/control the aforementioned parallelization modes. In contrast, the
new page has little of the former structure and it seems to tend to
explain command line options and how the options should be used
(rather than the parallelization modes and how these should be used).
The end result may be more "dummy-user" friendly (I mean novice ,
but I think it removes a lot of the "educational" value of the former.
Again, naturally, I'm biased towards the original structure and
content (as well as towards not treating users as dummies), but a page
with rich content pointing to external pages etc. is perhaps not the
right stuff for the version manual.
Along the lines of what I proposed a few weeks ago, the GPU
acceleration should perhaps get a separate page on the new website. It
would be the ideal place to port over the old gromacs.org/gpu content,
refresh the illustrations. I have some and will make new versions of
those that include dynamic pruning and PME (unless someone else wants
to . We should also update the performance numbers on that page --
and BTW, let's not loose the benchmark systems we have there as these
are the only "official" GROMACS benchmark inputs that people have
access to and do work (the public PRACE benchmarks are still
This page should not be versioned, I think.
#2 Updated by Mark Abraham 10 months ago
Szilard, if you think things are missing or can be expressed better, please write/update some words. It's not everybody else's job to make things different in a way that you prefer. The content that I imported and restructured from the wiki in 2015 is still substantially unchanged, so it seems everybody has voted with their feet. There are still TODOs regarding PME-on-GPU documentation that would also be good to clarify.
We modified the original GROMACS use of "acceleration" years ago, when we dropped GMX_CPU_ACCELERATION in favour of GMX_SIMD.