search

radical-cybertools / htbac

High throughput binding affinity calculator
MIT License
2 stars 1 forks source link

Specific NAMD compilation on Titan #17

Open jdakka opened 6 years ago

jdakka commented 6 years ago
mturilli commented 6 years ago

What is the requirement we are not satisfying? Having a MPI version of NAMD on Titan?

jdakka commented 6 years ago

@dww100 could you verify if this is what you meant by your original request?

dww100 commented 6 years ago

@jdakka can you send me what I originally said?

I can imagine several things this might refer to but I can't remember which specifically this was about:

  1. Can we get a version of NAMD running that will use multiple nodes with multiple GPUs (this may not actually be sensible performance wise)
  2. If we want to run TIES where we have to only use the CPUs then can we run on multiple nodes - realistically one will be very slow.
  3. I think (but am in no way sure) that in order to support the multiple replicas in one NAMD instance mentioned in #16 we need an MPI supporting version
jdakka commented 6 years ago

@dww100 the requirement from your email is as stated in the title Specific NAMD compilation on Titan?

  1. You'll have to check the software link that I included for a NAMD that is specific to your requirements. To use ORTE and achieve the 3750 nodes requirement, it is best to use a non-MPI version of NAMD. I checked that link, and there is a multicore-CUDA non-MPI version available.

  2. You'll have to check the software link that I included for a NAMD that is specific to your requirements. To use ORTE and achieve the 3750 nodes requirement, it is best to use a non-MPI version of NAMD. In order to run TIES, you need a non-CUDA, non-MPI multicore binary, which is also available in the link.

  3. In our past runs with you, we generally mapped a single replica to a NAMD executable. I see the frustration that this would lead in order to reach the 3750 nodes to hit the desired wall time, but I see this as a new requirement that would require some digging to see which NAMD binary would allow for multiple replicas to run in a single NAMD executable.