I also have to agree with Ross' comments here - the biggest problem with
really huge systems is that the sampling requirements are really huge too.
I am of course interested in the technical challenges of doing larger
simulations in the software itself, but I would not jump an order of
magnitude in my system size without expecting significant problems (I have
run up to 800,000-900,000 atoms, but I was looking at pme boxsize effects).
Regards - Bob Duke
----- Original Message -----
From: "Ross Walker" <ross.rosswalker.co.uk>
To: "'AMBER Mailing List'" <amber.ambermd.org>
Sent: Wednesday, June 01, 2011 10:16 AM
Subject: Re: [AMBER] RAM requirements
> Hi Dmitry,
>
>> I've tried to make a solvated protein in tleap with the command
>> "solvatebox M TIP3PBOX 275" to start with rather simple test system. It
>> used 8 Gb RAM, then 7.5 Gb swap and hanged. It probably means that
>> solvatebox routine memory usage is not optimal.
>
> Yes I was going to email about this as I tried it over the last few days.
> I
> got a 10 million atom system built in Leap but it had a malloc failure
> attempting to save the prmtop. So if you want to start simulating 10
> million
> atom systems plus you will need to do a bunch of work first before even
> getting to running the simulation. Nobody in the AMBER community that I
> know
> of simulates atoms much over 1 million atoms or so. This is mostly because
> at sizes above this you get such woefully poor sampling, with any MD code,
> that you really can't give anything more than anecdotal results.
>
> You could try sleap - maybe that does things better. Alternatively I'd add
> updating Leap to the list of work needed to run your 10 million atom
> system.
>
>> Interestingly, GROMACS successfully built this system using only 6 Gb
>> RAM, but I'm not familiar with that program and its forcefields. The
>> resulting PDB file size is 1 Gb. Is it possible to convert it into
>> AMBER format or I need to solvate it from scratch?
>
> There may be some tools available to convert gromacs files to amber prmtop
> and inpcrd although I have not tried. If you have the pdb now why not read
> this into leap and use setbox to give you the box size and then see what
> happens.
>
> As for memory usage in pmemd I have a 2 million atom test case and this
> uses
>
>
> Dynamic Memory, Types Used:
> | Reals 63553958
> | Integers 69171816
>
> | Nonbonded Pairs Initial Allocation: 58293216
>
> | Running AMBER/MPI version on 8 nodes
>
> on 8 nodes. So 748 MB per MPI task. So you'd probably need between 5 and
> 10
> times this for a 10 million atom system so you are probably looking at
> around 5 to 8GB per mpi task. So on an 8 way node you'd need around 64GB
> of
> ram which is certainly reasonable. Note though that the memory usage will
> actually go down some as you increase the number of MPI tasks so this
> would
> probably end up around 2GB a core (a guess!) on 128 cores.
>
> To be honest though most of the tools in the AMBER tool chain. Leap, pmemd
> and then ptraj would all benefit from being tuned specifically for such
> large simulations so if you plan to run these regularly it would probably
> be
> in your long term benefit to look at the codes and make some of these
> improvements. One example would be implementing a way to read the prmtop
> and
> coordinate file in parallel for example to avoid the large memory spike
> needed on the master thread etc.
>
> Good luck,
> Ross
>
> /\
> \/
> |\oss Walker
>
> ---------------------------------------------------------
> | Assistant Research Professor |
> | San Diego Supercomputer Center |
> | Adjunct Assistant Professor |
> | Dept. of Chemistry and Biochemistry |
> | University of California San Diego |
> | NVIDIA Fellow |
> | http://www.rosswalker.co.uk | http://www.wmd-lab.org/ |
> | Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
> ---------------------------------------------------------
>
> 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.
>
>
>
>
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
>
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Jun 01 2011 - 10:30:05 PDT
