[gmx-users] GROMACS on Mac OS X

[Erik Lindahl]

"Donald E. Elmore" wrote:

> Hi all--
>
> Our research group is planning on purchasing a new workstation, and were
> considering one of the new dual processor Mac G4 machines.  I remember
> people discussing running GROMACS under Mac OS X awhile ago, and I
> wondered what kinds of experiences they have had with that.  Have there
> been any problems with compiling or special considerations/changes people have
> needed to make?  And has anyone done any benchmarks with GROMACS on that
> platform?  We are also thinking about just going with another
> Pentium/Linux system like we presently use with GROMACS--any opinions on
> this?
>
> Thanks for your help!
>
> -Don Elmore

Hi Don,

Gromacs works just splendid under OS X, and a couple of months ago I even wrote
inner loops using Altivec ("velocity engine") loops for G4 and higher processors.
It's
a really nice platform that allows you to mix unix with applications like
powerpoint :-)

Nevertheless, you should not expect the performance to even approach that of
x86-based
clusters. The altivec velocity engine is extremely powerful for algorithms that
are perfectly
vectorizable (like photoshop filters and video editing), but molecular dynamics
is not one
of them.

We are increasingly starting to use Powerbooks and OS X for laptops in my group
at stanford,
since it's wonderful to be able to combine productivity applications with emacs
and other unix
stuff, but when it comes to clusters we rely on dual athlons.

The performance you get from Gromacs is roughly proportional to the specFP rating
since we
have assembly loops both for Athlon and PowerPC. Apple doesn't like to publish
specFP
numbers, but there are independent specFP95 benchmarks.  For each processor this
is:

Athlon XP 2000+ (1.7 GHz):    642 specfp2000 =(approx) 59.4 specfp95
PowerPC g4 733 MHz:  23.9 specfp95
PowerPC g4 1 GHz (estimated, scaled): 32.6 specfp95

In other words, the Atlon CPU is 82% faster for general floating-point problems.
Both processors
of course benefit significantly from the 3DNow/SSE and Altivec loops,
respectively.

Officially you should use MP instead of XP processors for SMP, but in practice
the XP works
just fine - the price of an SMP configuration with 2*1.7 GHz athlon is about
$1500. The dual G4
is considerably more expensive since you get a good graphics card, firewire,
superdrive and high
quality case for each machine.

Please note that I am NOT saying  you shouldn't order G4's! That question is much
more complicated.
While the Athlon wins on brute force performance and specfp/$ the G4 is probably
cheaper to
maintain, easier to install and wonderful to work with (you won't have to compile
your own linux
kernels...).

I've got access to a cluster of slightly older G4's, so I'll try to update the
benchmark page with
hard numbers for 3.1 this week.

Cheers,

Erik