Hi Francesco,
For 'performance' you need at least 1 GPU per replica when running REMD on
GPUs, it doesn't matter if these are in the same node, spread between
nodes or a combination of the two providing your mpirun is setup to
distribute threads correctly to different nodes. However, if you just want
to test / debug REMD on GPUs you can overload a GPU, performance will suck
but it will work. Just make sure the GPUs are NOT in compute exclusive
mode.
nvidia-smi -c 0
Then just run the number of MPI threads to match the number of groups you
have and they will get handed out in a round robbin fashion to the number
of GPUs that are visible to those MPI threads, overloading as need be.
All the best
Ross
On 9/10/13 11:03 PM, "Francesco Pietra" <chiendarret.gmail.com> wrote:
>Hello Hai:
>Can a debug trial be carried out on a shared memory GPU machine (the one I
>use currently) or are multiple physical nodes required? With NAMD,
>physical
>nodes are needed.
>
>At any event, what you suggest is MD, i.e., at a single temperature. So
>that the milliseconds are needed to hope to overcome the energy barriers.
>
>Thanks
>francesco
>
>
>On Tue, Sep 10, 2013 at 11:59 PM, Hai Nguyen <nhai.qn.gmail.com> wrote:
>
>> hi Francesco,
>>
>> If you can access to GPU cluster, you might want to try implicit
>> solvent simulation (like igb=8) first. Instead of running expensive
>> REMD, you can run multiple MD simulations (with different ig number or
>> different starting structures). With GPU cluster and igb8, you might
>> get about 500 ns/day.
>>
>> Hai
>>
>>
>> On Tue, Sep 10, 2013 at 11:22 AM, Francesco Pietra
>> <chiendarret.gmail.com> wrote:
>> >
>> > I was at your earlier papers. May I ask for a pdf of the 2013 cited
>> > article? At our poor academy we do not subscribe that journal <
>> > francesco.pietra.accademialucchese.it>.
>> >
>> > My aim was to hopefully arrive at a major cluster of conformations,
>>if
>> > any. It would be an improvement anyway on lack of useful diffraction
>>by
>> > this part of the protein (Following Modeller integration of lacking
>> > portions, I have detached a portion of the six X-ray diffr
>>indescribable
>> > portions, planning to superimpose the T-remd-resulting cluster through
>> the
>> > few well described helical turns and delete the original). NMR - in my
>> > hands - proved unable to treat the case. So, theory was left.
>> >
>> > It deals of a transmembrane protein, with the portions to model just
>> > emerging from the bilayer. Nonetheless, I could also try implicit.
>> >
>> > Thanks a lot
>> > francesco pietra
>> >
>> >
>> > On Tue, Sep 10, 2013 at 4:13 PM, Carlos Simmerling <
>> > carlos.simmerling.gmail.com> wrote:
>> >
>> > > I tend to agree- to my knowledge, this has never been carried out to
>> > > reasonable convergence in the literature for a peptide this size in
>> > > explicit water. By "convergence" I mean showing that very similar
>> results
>> > > are obtained regardless of starting conformation. That's stricter
>>than
>> some
>> > > reports where it is claimed that the populations are no longer
>> changing,
>> > > which I find not as satisfying as it should be (in my opinion).
>>Niel is
>> > > right, if you want to just do a conformational search, this could be
>> fine,
>> > > but if you're looking for the populations themselves to be reliable,
>> it's
>> > > still perhaps intractable.
>> > >
>> > > I'm not sure which comments of mine on implicit solvent you
>>reference.
>> our
>> > > GB models have gotten better, and with igb=8 you might find that you
>> get
>> > > what you need. Look at our article on it and see if it's what you
>>need,
>> > > though the test systems are much smaller than what you're proposing.
>> > >
>> > > 1. Nguyen, H., Roe, D. and Simmerling, C., ³Improved Generalized
>> Born
>> > > Solvent Model Parameters for Protein Simulations², J. Chem.
>>Theory &
>> > > Comput., 9, 20202034 (2013) DOI: 10.1021/ct3010485
>> > >
>> > >
>> > >
>> > > On Tue, Sep 10, 2013 at 10:01 AM, Niel Henriksen
>><shireham.gmail.com>
>> > > wrote:
>> > >
>> > > > Francesco,
>> > > >
>> > > > More than 200 replicas could be a major challenge, especially with
>> > > explicit
>> > > > solvent. If you hope to find some plausible folded
>>conformations, it
>> > > might
>> > > > work (i don't know that you need to go to 600K though). If you
>>hope
>> to
>> > > > converge the entire ensemble, you might want to avoid this
>>particular
>> > > > route. I had trouble converging a tetranucleotide in explicit
>>water
>> with
>> > > > ~4 microseconds/replica of simulation (24 replicas).
>> > > > http://www.ncbi.nlm.nih.gov/pubmed/23477537
>> > > >
>> > > > --Niel
>> > > >
>> > > >
>> > > > On Tue, Sep 10, 2013 at 6:45 AM, Francesco Pietra <
>> chiendarret.gmail.com
>> > > > >wrote:
>> > > >
>> > > > > >Assuming that I am interpreting your question correctly,
>> > > > >
>> > > > > Yes, and the answer is very useful: I can assume a larger delta
>>T
>> than
>> > > > the
>> > > > > 0.8K I reported. Provided that I will be able to calculate the
>> exchange
>> > > > > frequency (so far unclear to me how; I can just change the
>>number
>> of
>> > > > > steps/run, I used 100, if increased to 1000 there is no more any
>> > > > exchange).
>> > > > >
>> > > > > At any event, are >200 replicas (as needed with my system for
>> > > 314-600K) a
>> > > > > reasonable prospect or one that is destined to result as
>> technically
>> > > > > unfeasible? The literature is abundant of variants on small
>> peptides.
>> > > > >
>> > > > > Finally, is exhaustive MD equilibration an absolute need or it
>>can
>> just
>> > > > > help? So far I have not insisted in the equilibration at 314K
>> > > > (rmsd/frame
>> > > > > is still raising because - I suppose - of the flexibility of the
>> > > portion
>> > > > to
>> > > > > model: clearly there are low energy barriers)
>> > > > >
>> > > > > Thanks
>> > > > >
>> > > > > francesco pietra
>> > > > >
>> > > > >
>> > > > > On Tue, Sep 10, 2013 at 1:12 PM, Jason Swails <
>> jason.swails.gmail.com
>> > > > > >wrote:
>> > > > >
>> > > > > > On Tue, Sep 10, 2013 at 3:38 AM, Francesco Pietra <
>> > > > chiendarret.gmail.com
>> > > > > > >wrote:
>> > > > > >
>> > > > > > > Hello:
>> > > > > > > I am progressing toward a major task of carrying out a pure
>> T-remd
>> > > > > > > (parallel tempering) with a 34aa peptide in explicit water
>>for
>> a
>> > > > > 314-600K
>> > > > > > > range. After C. Simmerling reports about implic media, I did
>> not
>> > > > > attempt
>> > > > > > > any simulation under implicit media.
>> > > > > > >
>> > > > > > > I started with 8 replicas and geom progressing temp,
>>noticing,
>> for
>> > > > the
>> > > > > > > 314-320K range, full exchange is obtained. deltaT is 0.8K.
>> > > > > > >
>> > > > > > > On increasing the nr of replicas and delta T on the above
>> basis,
>> > > thus
>> > > > > 16
>> > > > > > > replicas for 314-326K, the exchange is no more so good. That
>> is,
>> > > the
>> > > > > > > highest and the lowest temp do not exchange with one another
>> > > > directly.
>> > > > > > >
>> > > > > > > My question is, do, for example, exchange of 16 with 15
>>only,
>> and
>> > > > > > exchange
>> > > > > > > of 14 with 15, compensate? In other words, do "local
>>exchanges"
>> > > > > > propagate?
>> > > > > > >
>> > > > > > > In this affair, I restrained the dihedrals for a two
>>(starting)
>> > > alpha
>> > > > > > > helical turns, which are well defined by X-ray diffraction.
>>The
>> > > rest
>> > > > of
>> > > > > > the
>> > > > > > > peptide does not clearly diffract at 100K, and I am
>>attempting
>> to
>> > > > > > simulate
>> > > > > > > how the conformation, or prevailing cluster of
>>conformations,
>> > > should
>> > > > > be.
>> > > > > > I
>> > > > > > > did not apply any other restraint to this undefined part.
>> Should
>> > > the
>> > > > > > > chirality at the peptide bond be imposed even if the
>>prevailing
>> > > > > > > conformation is unknown?
>> > > > > > >
>> > > > > > > Thanks so much for advice
>> > > > > > >
>> > > > > >
>> > > > > > Assuming that I am interpreting your question correctly, there
>> is no
>> > > > need
>> > > > > > to achieve exchanges between the first and last temperature
>> replica.
>> > > > In
>> > > > > > fact, if you are achieving exchanges between those two a high
>> > > > percentage
>> > > > > of
>> > > > > > the time, then all of the replicas in between are unnecessary
>> :). As
>> > > > > long
>> > > > > > as adjacent replicas exchange frequently enough (about 20% of
>>the
>> > > time
>> > > > is
>> > > > > > good in my experience), then you will get good enough mixing
>>in
>> state
>> > > > > space
>> > > > > > to make efficient use of REMD. (Replicas will be able to
>>diffuse
>> > > > through
>> > > > > > state space from the lowest temperature to the highest
>>_through_
>> the
>> > > > > > intermediate temperatures).
>> > > > > >
>> > > > > > The following website will give you a good set of starting
>> > > temperatures
>> > > > > > between two replicas: http://folding.bmc.uu.se/remd/ (the NPT
>> > > results
>> > > > > are
>> > > > > > good enough for the NVT REMD that Amber runs). It may be
>>worth
>> > > > running a
>> > > > > > quick simulation to verify that all exchange rates are
>> approximately
>> > > > what
>> > > > > > you want them to be before starting major production.
>> > > > > >
>> > > > > > HTH,
>> > > > > > Jason
>> > > > > >
>> > > > > > --
>> > > > > > Jason M. Swails
>> > > > > > BioMaPS,
>> > > > > > Rutgers University
>> > > > > > Postdoctoral Researcher
>> > > > > > _______________________________________________
>> > > > > > 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
>> > > > >
>> > > > _______________________________________________
>> > > > 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
>> > >
>> > _______________________________________________
>> > 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
>>
>_______________________________________________
>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 Sep 11 2013 - 07:00:04 PDT
