Re: [Apbs-users] Submitting async parallel jobs

[Ryogo S. <rsu...@gm...>]

Hi Nathan,

Thanks for your response with list of literatures.
Actually, i've read the first two before, but still wasn't sure
how communication between subdivisions was carried out during parallel
calculation.
So, the final answer to this question is
there's no communications between subdivisions during the calculation
except at the final processing step even in mpi version of mg-para, correct=
?

It would be nice for future update if we can perform multiple
iteration of such processes,
(by reusing potentials and bondary conditions of the subdivision from previ=
ous
iteration after equilibration process) so that it will be truly
equivalent with running
a calculation with a single huge grid and be less dependent with ofrac valu=
es.
(Then we can divide the region as much as we want (up to the capacity
of HDD of course...)
without any significant errors due to the subdivisions.)
Of course, that's a lot of work, so I don't expect it to be realized
any time soon,
but that would be a major update of features many people would be
interested in (at least I.)

Meanwhile, I'll test the current parallel calculation scheme with my own sy=
stem
and satisfy with it if the error seems to be reasonably small,
or try to devise some work-around if inter-division-communication is requir=
ed.

Thank you,

Ryogo


2005/10/25, Nathan Baker <ba...@bi...>:
> Hi Ryogo --
>
> You are correct, we make the assumption that values in the overlap
> region are sufficient to provide accurate boundary conditions.  This
> is the same principle as mg-para and is discussed in more detail in:
>
> Baker NA, Sept D, Holst MJ, McCammon JA. The adaptive multilevel
> finite element solution of the Poisson-Boltzmann equation on massively
> parallel computers . IBM Journal of Research and Development 45,
> 427-438, 2001. (http://www.research.ibm.com/journal/rd/453/baker.pdf)
>
> and
>
> Baker NA, Sept D, Joseph S, Holst MJ, McCammon JA. Electrostatics of
> nanosystems: application to microtubules and the ribosome. Proc. Natl.
> Acad. Sci. USA 98,  10037-10041, 2001.
> (http://dx.doi.org/10.1073/pnas.181342398)
>
> The concept is the same as the traditional focusing method:
>
> Gilson MK, Honig BH.  Calculation of electrostatic potentials in an
> enzyme active site.  Nature.  330 (6143), 84-6, 1987.
> http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=3DRetrieve&db=3DPubMed&=
list_uids=3D3313058&dopt=3DAbstract
>
> but the notion of extending this to parallel calculations (and the
> resulting asynchronous generalization) is based on work by Mike Holst
> and Randy Bank:
>
> Bank RE, Holst M.  A new paradigm for parallel adaptive meshing
> algorithms.  SIAM Review.  45 (2), 291-323, 2003.
> http://dx.doi.org/10.1137/S003614450342061
>
> In short, the asynchronous mode does the same operations as mg-para
> but leaves out the final step of reassembling to the user.
>
> Hope this helps,
>
> Nathan
>
> On 10/25/05, Ryogo Sugitani <rsu...@gm...> wrote:
> > Hi Todd,
> >
> > Thanks for your reply.
> > But the main question still remains.
> > How is async parallel calculatiion different from running
> > separate serial calculations on sub-divided region and merge them
> > together afterwards?
> > In other words, how does the values on the border of one region is
> > transmitted to
> > the other regions without any integration master process?
> > (I assume mergedx is just really merging dx files and don't do much
> > equilibration calculcation or anything.)
> > Do you assume the values in overlap region would give a reasonable esti=
mates
> > of adjacent regions so that series of async parallel calculation is equ=
ivalent
> > with running a big single PB calculation?
> > is that the same principle used in mg-para option, too?
> >
> > If my questions is still unclear, please let me know.
> >
> > Thank you,
> >
> > Ryogo
> >
> > 2005/10/25, Todd Dolinsky <to...@cc...>:
> > > Hi Ryogo -
> > >
> > > The asynchronous use of APBS does just that - allows users to break u=
p a
> > > parallel calculation one processor at a time, and then asynchronously
> > > (one after another) run APBS.
> > >
> > > So if you have a large PQR file, for example, and it is recommended t=
hat
> > > you use parallel processing yet you only have one processor, you woul=
d
> > > split the file up (as you did), and then proceed to call APBS for eac=
h
> > > file, just as you mentioned:
> > >
> > > apbs apbs-PARALLEL-PE0.in
> > > apbs apbs-PARALLEL-PE1.in
> > > apbs apbs-PARALLEL-PE2.in
> > > apbs apbs-PARALLEL-PE3.in
> > >
> > > This allows users to run APBS on large PQR files with machines with
> > > limited resources.  Depending on what you want to get back from APBS =
you
> > > can either sum energies, or merge DX files using the mergedx utility.
> > > The naming of the files assists with the mergedx utility.
> > >
> > > On the other hand, if you have a multiple-processor machine and want =
to
> > > run APBS in parallel, there is no need to run APBS asynchronously - y=
ou
> > > just need to use the mg-para keyword.
> > >
> > > Please see
> > >
> > > http://apbs.sourceforge.net/doc/user-guide/x645.html#mg-para
> > > http://apbs.sourceforge.net/doc/user-guide/x645.html#async
> > >
> > > for more information.
> > >
> > > Todd
> > >
> > >
> > > Ryogo Sugitani wrote:
> > > > Hello,
> > > >
> > > > After you get appropriate sets of input file from a mg-para
> > > > calculation input file
> > > > with appropriate "async {rank}" options (by using asynchronous.sh o=
r by hand),
> > > > how do we actually submit those sets of input files into ABPS?
> > > >
> > > > For example, with born example, I've got
> > > > apbs-PARALLEL-PE0.in
> > > > apbs-PARALLEL-PE1.in
> > > > apbs-PARALLEL-PE2.in
> > > > apbs-PARALLEL-PE3.in
> > > > after I process apbs-PARALLEL.in with asynchronous.sh
> > > >
> > > > Do we just do feed the first input file, like
> > > > apbs apbs-PARALLEL-PE0.in
> > > >
> > > > or some other way?
> > > > (In case you calculate them individually, how would each division c=
an
> > > > communicate with each other?)
> > > >
> > > > I'll test the first case later today, but I don't see how apbs will
> > > > recognize which files to read
> > > > after reading the first one. Is the naming of the files the key?
> > > >
> > > > Thanks in advance,
> > > >
> > > > Ryogo
> > > >
> > > >
> > > > -------------------------------------------------------
> > > > This SF.Net email is sponsored by the JBoss Inc.
> > > > Get Certified Today * Register for a JBoss Training Course
> > > > Free Certification Exam for All Training Attendees Through End of 2=
005
> > > > Visit http://www.jboss.com/services/certification for more informat=
ion
> > > > _______________________________________________
> > > > apbs-users mailing list
> > > > apb...@li...
> > > > https://lists.sourceforge.net/lists/listinfo/apbs-users
> > > >
> > > >
> > >
> >
> >
> > -------------------------------------------------------
> > This SF.Net email is sponsored by the JBoss Inc.
> > Get Certified Today * Register for a JBoss Training Course
> > Free Certification Exam for All Training Attendees Through End of 2005
> > Visit http://www.jboss.com/services/certification for more information
> > _______________________________________________
> > apbs-users mailing list
> > apb...@li...
> > https://lists.sourceforge.net/lists/listinfo/apbs-users
> >
> >
>
>
> --
> Assistant Professor, Dept. of Biochemistry and Molecular Biophysics
> Center for Computational Biology, Washington University in St. Louis
> Web: http://cholla.wustl.edu/
>