[Beowulf] Thought that this might be of interest

[Joshua Baker-LePain]

On Mon, 6 Nov 2006 at 8:57am, Jeffrey B. Layton wrote

> I'm also not sure if LS-Dyna is built with the Intel compilers or
> not. This will also have an impact on performance since you have
> to _dumb-down_ the compile options to get it to run on Opteron
> (I mentioned this to Joe Landman, but for everyone's edification,
> here's a link:

They (LSTC) release binaries of each point revision compiled with both the 
PGI and the Intel compilers and name the binaries appropriately (I tested 
with both versions on both chips).  For the MPI binaries, they make this 
explicit (see Table 3 in the PDF I linked to).  For the SMP binaries, they 
call them 'amd' and 'xeon' (Tables 1 and 2).  The README states that the 
compiler for the 'amd/PGI' versions is PGI 6.1 and for the 'xeon/Intel' 
version it's ifort 8.1.

> I haven't seen the Matlab results before. Those are interesting.
> Can you explain the benchmarks a bit more? Is it the stock Matlab

I'll quote you what the grad student told me about the first one (Table 
4), which was actually written by his undergrad:

\begin{quote}
It appears to be reading in FEM nodal data, then selecting a subset region
to work with based upon user input. Nodal displacement data is
(heavily) interpolated through time, then a subset time span is chosen. A
(large) number of scatterers are then placed randomly inside the subset
region. The delaunay transformation (computationally expensive) is used to
find scatterers that would fall within the lumen of the heart
chamber, then remove them. I don't know much about this step aside from
the fact that it uses convex hulls. Once all the scatteres are in place,
a trilinear interpolation is used to find the displacement of each
scatterer based upon the motions of the nodal coordinates around it.
\end{quote}

Like I say in the table caption, that benchmark generates a large amount 
of system load -- it's about a 70/30 user/system split.

Table 5 benchmarks Field II, which is an ultrasound simulation library for 
Matlab <http://www.es.oersted.dtu.dk/staff/jaj/field/>.  It's distributed 
as a .mex file compiled *long* ago.  So that benchmark is more about Field 
than about Matlab -- but it's relevant for us.

> or did you substitute the BLAS library? That might make a
> difference.

That was stock Matlab R14p3.  I'm in the process of getting an updated 
license file for the most recent version, and then maybe I'll look at 
playing with the BLAS library.

-- 
Joshua Baker-LePain
Department of Biomedical Engineering
Duke University