[gmx-users] Numerical Stability of Gromacs Implementation

Mark Abraham mark.j.abraham at gmail.com
Fri Aug 22 17:05:36 CEST 2014


On Aug 22, 2014 4:49 PM, "Johnny Lu" <johnny.lu128 at gmail.com> wrote:
>
> it has many (relatively) power7 cores and aix 6.1.0.0 with no gpu, and I
> compiled single precision gromacs 4.6.6 with open mp on it.
> I tried xlc or mpcc for 30+ times.

6.1 was released 7 years ago, but if you have a functional xlc, then you
should be able to build GROMACS. It won't run fast, though.

>The gcc on it doesn't support openmp,
> and I compiled another one with openmp, and also gnu cmake.
> Linker is aix ld, with -lmass, and -lm flags. I am not sure if the mass
> library helped in compilation of fftw3 library.

It doesn't.

> Somehow, gcc can't find affinity support on aix ld, so I specified that in
> the job script of llqueue.
> May be next time I would try the mass simd library(-lmass_simdp7) or
vector
> library (-lmassvp4)
> (http://www-01.ibm.com/support/docview.wss?uid=swg27005375)
>
> Somehow, 24 intel xeon cpu (compiled with intel compiler, and mkl library)
> seems 4 times faster than 32 power7, when I run gromacs 4.6.6.

Not surprising, that Intel hardware has been the target of a lot of
optimization.

Mark

> On Thu, Aug 21, 2014 at 7:28 PM, Mark Abraham <mark.j.abraham at gmail.com>
> wrote:
>
> > On Thu, Aug 21, 2014 at 7:59 PM, Johnny Lu <johnny.lu128 at gmail.com>
wrote:
> >
> > > Sorry for asking this. Is it possible for me to get some references
about
> > > the claims of instability of gromacs simulations, and their debunk?
> > >
> >
> > Not really. You can find a paper from some of the Desmond authors that
> > correctly observes some issues in GROMACS 3.x. People occasionally
refer to
> > it here as if it is current news. These issues are long fixed, but not
> > worth writing about - journal articles should be about delivering
quality
> > science. More commonly, an issue would be handled via private email,
though
> > even these are rare. Overall, the biomolecular MD community is quite
good
> > at finding problems with their own and each other's algorithms and
> > implementations and getting them fixed constructively. People saying
> > "someone said x was bad but gave me no details" need to talk to the
> > someone, not just the authors of x ;-) If people start being evasive or
> > secretive about possible problems with their code... be concerned.
> >
> > From the few papers that I read, I guess algorithms of molecular
dynamics
> > > do not treat all observables equally well.
> > > Some old papers say that the velocity in velocity verlet is not
> > symplectic,
> > > but rather follows some shadow hamiltonian or generalized
equipartition
> >
> > theorem.
> > >
> >
> > This is common to all methods with a finite time step - see
> > http://en.wikipedia.org/wiki/Energy_drift. There are certainly a variety
> > of
> > ways of estimating the velocity with common integrators, and they have
> > different quality attributes. You can read about how GROMACS handles
this
> > in the manual.
> >
> > Then another one mentioned force splitting can reduce the resonance
effect
> > > caused by integrator.
> > > That said, very few papers talk about this.
> > >
> >
> > There are lots of papers that discuss details of multiple time-stepping
> > algorithms that seek to deal with this issue directly.
> >
> >  I don't know much about the effect of MD on the observables that I try
to
> > > look at.
> > >
> >
> > It's not an easy topic - generating converged sampling to assess
whether an
> > integration scheme correctly samples a complex observable is still a
> > non-trivial matter. That needs to happen before questions of "how much
> > algorithmic energy drift is acceptable" can be satisfactorily addressed.
> > Until then, claims of "my energy conservation is better than yours"
need to
> > be considered alongside "my number of independent converged-ensemble
> > samples is better than yours."
> >
> >  And, sorry for replying this late, I have been installing gromacs on
aix
> > > for a week.
> > > Compiling gcc took 3 days of computer time.
> > >
> >
> > Seriously, don't bother. I don't think there is any system that would
have
> > AIX, with no gcc package available, and which GROMACS 5.0 would run
> > decently on (which would require SIMD support, which currently means
x86 or
> > BlueGene/Q). I'd guess your laptop will get equivalent performance.
> >
> > Mark
> >
> >
> > >
> > > On Thu, Aug 14, 2014 at 2:29 PM, Mark Abraham <
mark.j.abraham at gmail.com>
> > > wrote:
> > >
> > > > On Wed, Aug 13, 2014 at 1:19 PM, Johnny Lu <johnny.lu128 at gmail.com>
> > > wrote:
> > > >
> > > > > Hi again.
> > > > >
> > > > > Some of my friends said that gromacs had lower numerical stability
> > than
> > > > > amber, and occasionally has mysterious error in energy.
> > > > >
> > > >
> > > > Show us a result and we'll discuss fixing it ;-)
> > > >
> > > >
> > > > > Is that still true? Does the implementation of Integrator cause
more
> > > > > resonance effect?
> > > > >
> > > >
> > > > Any numerical software can be used poorly. Any numerical software
can
> > > have
> > > > bugs. Give them the same input to two implementations of the same
> > > algorithm
> > > > and they should give results that are similar (whatever that means
in
> > the
> > > > problem domain).
> > > >
> > > >
> > > > > I am trying to run NVE simulation with the single precision
version,
> > > and
> > > > so
> > > > > I guess the implementation of thermostats is not my most immediate
> > > > concern
> > > > > at the moment.
> > > > >
> > > > > I searched online, and haven't found much about numerical
instability
> > > of
> > > > > gromacs using amber forcefield (I mean for version 4.6, which
> > supports
> > > > the
> > > > > amber force-fields without modification.)
> > > > >
> > > >
> > > > What do you mean by numerical stability? Should you actually be
> > concerned
> > > > about getting a correct physical observable? People look at energy
> > > > conservation because it is easy to do, but connecting that to the
> > quality
> > > > of the resulting physics is non-trivial. Even if you can (say)
deliver
> > > > state-of-the-art energy conservation (and most packages can do this
in
> > > some
> > > > mode), you still need to measure the price in simulation throughput
> > > > incurred by running in that mode. The overall objective is to
deliver
> > > > science of sufficient quality with minimal time to solution. In some
> > > cases,
> > > > it is right to let some energy leak and let the thermostat fix it,
if
> > you
> > > > generate many more samples by so doing.
> > > >
> > > > Why do people still use and paid for amber, when gromacs is free?
> > > > > I look online and the number of papers that use amber is about
twice
> > of
> > > > the
> > > > > number of paper that use gromacs. There are very few papers that
> > > > mentioned
> > > > > both gromacs and amber.
> > > > >
> > > >
> > > > Inertia is a big effect. You tend to adopt the package used by the
> > local
> > > MD
> > > > expert so you have a low-latency person of whom to ask questions.
They
> > > tend
> > > > to still use the package they used in their doctoral work,
regardless
> > of
> > > > current technical merit, because learning to use such software well
is
> > > > tricky and discarding your old workflows is not a cheap proposition.
> > The
> > > > next biggest effect is what functionality is available, and at what
> > > > performance.
> > > >
> > > > Mark
> > > >
> > > > Thank you.
> > > > > --
> > > > > Gromacs Users mailing list
> > > > >
> > > > > * Please search the archive at
> > > > > http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> > > > > posting!
> > > > >
> > > > > * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> > > > >
> > > > > * For (un)subscribe requests visit
> > > > > https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users
> > or
> > > > > send a mail to gmx-users-request at gromacs.org.
> > > > >
> > > > --
> > > > Gromacs Users mailing list
> > > >
> > > > * Please search the archive at
> > > > http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> > > > posting!
> > > >
> > > > * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> > > >
> > > > * For (un)subscribe requests visit
> > > > https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users
or
> > > > send a mail to gmx-users-request at gromacs.org.
> > > >
> > > --
> > > Gromacs Users mailing list
> > >
> > > * Please search the archive at
> > > http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> > > posting!
> > >
> > > * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> > >
> > > * For (un)subscribe requests visit
> > > https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> > > send a mail to gmx-users-request at gromacs.org.
> > >
> > --
> > Gromacs Users mailing list
> >
> > * Please search the archive at
> > http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> > posting!
> >
> > * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> >
> > * For (un)subscribe requests visit
> > https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> > send a mail to gmx-users-request at gromacs.org.
> >
> --
> Gromacs Users mailing list
>
> * Please search the archive at
http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
send a mail to gmx-users-request at gromacs.org.


More information about the gromacs.org_gmx-users mailing list