Re: [AMBER] Regarding the MPI error in parallel running of 3D-RISM

From: Tyler Luchko (Lists) <"Tyler>
Date: Fri, 6 Jul 2018 16:17:48 -0700

> On Jul 5, 2018, at 3:39 AM, PRITI ROY <> wrote:
> Dear David,
> I am doing 3D-RISM with the following parameters:
> ----buffer 15.0
> --grdspc 0.5,0.5,0.5
> --closure KH
> --centering -1
> Before using 48 core, tried with fewer cores and every time getting the
> error "FATAL: allocation failure in ivector()" until the reaching 48 cores.
>> From the Amber mailing list, I came to know this error come due to memory
> issue.

How much RAM does your machine have? A system of this size should be < 10GB.

3D-RISM is a distributed memory program, so you can use MPI across multiple computers to increase the aggregate memory available. However, a 48 core computer should have more than enough memory for a system like this. It would be helpful to know more about the solvent you are using and the hardware.

> I also tried with a system of 2775 atoms in different combination of cores
> and time taken in every combination cores to run of five frames is as
> follows:
> No. of Cores Time (minutes)
> 6 10
> 8 8
> 10 8
> 12 5
> 24 5
> I could not understand why my system taking longer time and 48 cores for
> only one frame.
> Do I need to change or add any parameter which can do the scaling up?

3D-RISM does not scale linearly with the system size or number of processors but I suspect that something else is going on. As was suggested, try running a non-MPI job to completion and use the --verbose 2 and --progress flags to observe how the calculation is converging. This will also give you information about the grid dimensions. If you don’t have enough memory run the calculation, then you need to find a different computer.

Once you get this running, I suggest that you also try larger buffers and smaller grid spacing to make sure you are getting the numerical precision you desire. I expect that you will need even more memory than what is currently required.

Hope this helps,


> On Thu, Jul 5, 2018 at 7:18 AM David A Case <> wrote:
>> On Thu, Jul 05, 2018, PRITI ROY wrote:
>>> Then another error was appeared which is as follows:
>>> "rism3d.snglpnt.mpi: error while loading shared libraries:
>>> cannot open shared object file: no such file or
>>> directory"
>>> and it is resolved by setting LD_LIBRARY_PATH as "export
>>> LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/../amber16/lib""
>> Note that this should be done in the $AMBERHOME/ script. You
>> should
>> expect to run this script everytime you log in -- most users put something
>> like this in their startup script:
>> export AMBERHOME=/path/to/amber18
>> test -f $AMBERHOME/ && source $AMBERHOME/
>>> Now I am stuck in memory problem. My system has 5550 atoms, running one
>>> frame of the trajectory with 48 cores and after 3hours its not yet
>>> finished. I have 300ns long trajectory. Is it possible to speed up this
>>> calculation with this system size?
>> Indeed, rism3d can be extremely time consuming for large systems. But
>> don't
>> assume that more cores are always better: have you tried your system with
>> fewer cores (say 4 or 8, or even 1)? If you use the --progress option, you
>> can see what is happening: you may need to tweak convergence properties.
>> It's hard to say more without knowing what closure, grid spacing, etc you
>> are
>> using. It's worth gaining experience on smaller systems that can complete
>> in
>> a few minutes, then scaling up.
>>> Can I run this 3d-RISM calculation in GPU as I couldn't found any GPU
>> based
>>> executable of rism3d.snglpnt?
>> We don't support 3D_RISM on GPU's (and as far as I know, no one else does
>> either).
>> ....good luck....dac
>> _______________________________________________
>> AMBER mailing list
> _______________________________________________
> AMBER mailing list

AMBER mailing list
Received on Fri Jul 06 2018 - 16:30:02 PDT
Custom Search