Support

Support Options

Submit a Support Ticket

 
  • Discoverability Visible
  • Join Policy Invite Only
  • Created 29 Jun 2010

Parallelisation of the schroedinger solver

  1. Ross Maspero

    Hi,

    I was juts wondering if the automatic distribution of k-points across mpi-processes has been implemented into the schroedinger solver yet because I have tried to run it in serveral different configurations and it only seems to be able to partition it spatially

    Regards

    Ross

    Report abuse

  2. Tillmann Christoph Kubis

    Hi, Schroedinger runs in parallel. All you have to do is to define your number of k-points and call nemo5 with mpiexec -n number_of_CPUs. Please let us know if that works for you. Thanks, Tillmann

    Report abuse

  3. Ross Maspero

    Hi,

    I think I have realised what was confusing me. I was using firstly using the opnempi instead of mpich2 and then when I used mpich2, it didnt look like it was working because it only output one of the processes to terminal

    Thank you

    Ross

    Report abuse

  4. Ross Maspero

    Hi,

    Is there any way to drastically reduce the memory requirement per process for the calculating the bulk band structure? I require a band structure of a system in excess of 4000 atoms however the memory requirement for 2000 atoms is over 20gb per process which makes it impossible to run more than 1 process per node.

    An example of an input file I use is below which is for a 686 atom system and when I include partitioning, it has no effect on the memory per process.

    (file:GaBiAs_parallel.in not found)

    Is there something I have done wrong in this input or is there better way to optimise the memory?

    Thank you

    Ross

    GaBiAs_parallel.in

    Report abuse

  5. James Fonseca

    Hi Ross, Where are you running this? your own cluster or through nanoHUB workspace? Thanks, Jim

    Report abuse

  6. Ross Maspero

    Hi, I am running it on my own cluster. Ross

    Report abuse

  7. James Fonseca

    Hi Ross, Do you need all the eigenvalues/eigenvectors or do you just want to see an Ek diagram?


    Emailed forum response from jefonseca@gmail.com

    Report abuse

  8. James Fonseca

    Hi Ross, Michael is going to look into this. Thanks for your patience. Jim

    Report abuse

  9. Ross Maspero

    Hi

    I need to eigenvalues and eigenvectors as well but it still has this memory requirement when I try to calculate just the Ek diagram

    Thank you

    Ross

    Report abuse

  10. Ross Maspero

    Hi,

    What do you mean by storage of eigenvectors? Do you mean that a single k-point’s eigenvectors for a 10×10×10 supercell adds up to 20gb and does this mean if I disable this then I will no longer have access to them in the program because I require the current eigenvectors of that iteration for unfolding the supercell band structure.

    Thank you

    Ross

    Report abuse

  11. Ross Maspero

    Hi,

    Thank you for all your help. Do you have an idea of the time needed to fix this and also would I be allowed to know how it was fixed so I can implement it in my version of Nemo5?

    Thank you

    Ross

    Report abuse

  12. Ross Maspero

    Hi,

    Has there been any progress with the memory problem?

    Thank you

    Ross

    Report abuse

  13. Ross Maspero

    Hi,

    Any news would be appreciated

    Ross

    Report abuse

  14. James Fonseca

    Hi Ross, 4,000 atoms would require 102GB to store everything (4,000*20 (sp3d5s*))^2 * 16 bytes = 102GB


    Some suggestions: calculate only desired eigenvalues and eigenvectors

    eigen_values_solver = krylovschur

    number_of_eigenvalues should be on the order of 10,20,100…

    also, it helps to specify a value for ‘shift’

    some of these are given in the quantum_dot/test_schroedinger.in (as well as other input files) in the regression test resource https://nanohub.org/resources/19171

    Let us know how that goes.

    Thanks, Jim

    Report abuse

  15. Ross Maspero

    Hi,

    Thank you for the suggestion. I will give that a try. What does “shift” actually do? Does it set some limit?

    Thank you

    Ross

    Report abuse

  16. Ross Maspero

    Hi,

    Just to let you know that the selecting the number of wanted eigenvalues works well for a unit cell band structure calculation but not for any of the supercell calculations I have done which is what I vitally need it for my project

    Ross

    Report abuse

  17. Ross Maspero

    Hi,

    I have added a shift value but this does not really matter too much at the moment because the band structure is incorrect. Can I confirm that selecting the wanted number of eigenvectors does not work for supercells at all in Nemo5 and that it is not just an error in my compiled version?

    Thank you

    Ross

    Report abuse


nanoHUB.org, a resource for nanoscience and nanotechnology, is supported by the National Science Foundation and other funding agencies. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.