Re: AMBER: Amber parallel version on CPU power of 2 only

From: Robert Duke <rduke.email.unc.edu>
Date: Tue, 29 Jan 2008 13:43:03 -0500

Fatima -
No problem with starting a run in sander and continuiing it in pmemd. There
are some limitations in terms of what pmemd supports, but it currently does
all the generic pme and generalized Born stuff (look at manual for list of
unimplemented capabilities, or simply try running it, and it will refuse to
do things that it can't do correctly). On mdin input, the biggest difference
is pmemd still only supports the old group-based ways of specifying
restrained or belly atoms; this is documented through amber 7 at least
(possibly also 8). PMEMD should be substantially faster though if you have
a decent interconnect for your cluster, and probably worth using even with
gigabit ethernet.
Regards - Bob Duke

----- Original Message -----
From: <fatima.chami.durham.ac.uk>
To: <amber.scripps.edu>
Sent: Tuesday, January 29, 2008 1:19 PM
Subject: RE: AMBER: Amber parallel version on CPU power of 2 only


> Dear Ross,
>
> So its an inherent restriction to Sander and not to the MPI
> implementation.
>
> For my system of 11500 sites the 16 CPUS run is my cutoff (2times faster
> than 1cpu)
>
> I am thinking to use PMEMD to run my MD simulation for a long length of
> time (20ns)
>
> are input and output files for PMEMD compatibles to the one that
> sander.mpi uses
> cause I will carry on the simulation and not start from scratch
>
>
> best wishes
> fatima
>
>
>
>
>
>
> Quoting Ross Walker <ross.rosswalker.co.uk>:
>
>>
>> Hi Fatima,
>>
>> Amber 9 should support non-power of 2 cpus for sander.MPI. However, the
>> efficiency will go down if you use non-powers of two since the code
>> switches
>> from using a bindary tree to a standard allreduce for communication of
>> forces. How much this hurts will depend on your MPI implementation. You
>> could for example find that say 12 cpus is actually slower than 16 cpus.
>>
>> In Amber 8 you need to recompile a special parallel binary that has
>> -DNOBTREE defined in config.h. Note, however that this will turn off the
>> binary tree for ALL processor counts, so 8 cpus with the NOBTREE code
>> will
>> run slower than 8 cpus with the regular code. (this is not the case with
>> amber 9)..
>>
>> Alternatively if you are using a method supported by PMEMD, regular MD /
>> GB
>> etc then I would suggest using PMEMD. It will give you better performance
>> and doesn't have the power of two processor limitation that sander does.
>> See
>> $AMBERHOME/src/pmemd.
>>
>> All the best
>> Ross
>>
>> /\
>> \/
>> |\oss Walker
>>
>> | Assistant Research Professor |
>> | San Diego Supercomputer Center |
>> | Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
>> | http://www.rosswalker.co.uk | PGP Key available on request |
>>
>> Note: Electronic Mail is not secure, has no guarantee of delivery, may
>> not
>> be read every day, and should not be used for urgent or sensitive issues.
>>
>> > -----Original Message-----
>> > From: owner-amber.scripps.edu
>> > [mailto:owner-amber.scripps.edu] On Behalf Of
>> > fatima.chami.durham.ac.uk
>> > Sent: Tuesday, January 29, 2008 05:39
>> > To: amber.scripps.edu
>> > Subject: AMBER: Amber parallel version on CPU power of 2 only
>> >
>> >
>> > Dear Amber's users,
>> >
>> > Sander.MPI runs on 2,4,8,16,32,64 but not on any other
>> > number of processors
>> > which is not a power of 2 !!
>> >
>> > is there a way to solve this restriction ?
>> >
>> >
>> > many thanks
>> >
>> > fatima
>> > --------------------------------------------------------------
>> > ---------
>> > The AMBER Mail Reflector
>> > To post, send mail to amber.scripps.edu
>> > To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>> >
>>
>>
>> -----------------------------------------------------------------------
>> The AMBER Mail Reflector
>> To post, send mail to amber.scripps.edu
>> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>>
>>
>
>
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber.scripps.edu
> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>


-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Wed Jan 30 2008 - 06:07:29 PST
Custom Search