FAQ

1. Why is my NWChem simulation running slowly?

Several factors could contribute to your NWChem simulation running slower than expected. One common reason is that the simulation might be too large for the machine you're using. If this is the case, try running the simulation on a more powerful machine or a different machine type. Additionally, optimizing your simulation setup can improve performance.

Common Pitfall: Not Setting the n_vcpus parameter

When n_vcpus is not explicitly set, NWChem defaults to using all available cores on the selected machine. This can cause performance issues because NWChem employs both OpenMPI and OpenMP for parallelization. If all the cores are utilized without proper tuning, you may overload the machine.

For example, if you're using a c2-standard-4 machine (which has 4 CPU cores), and you don't specify the n_vcpus parameter, NWChem will:

• Launch 4 MPI processes (1 per core)
• Each MPI process will spawn 1 OpenMP thread

This results in 4 processes + 4 thread = 8 computation units—which is not ideal for a 4-core machine.

Recommended Solution

To avoid overloading your machine, explicitly limit the number of cores, like so:

task = nwchem.run(
    input_dir="/Path/to/SimulationFiles",
    sim_config_filename="sim.nw",
    n_vcpus=2,
    on=cloud_machine)

Tip: A good rule of thumb is to start with n_vcpus set to half the number of cores on your machine. From there, you can experiment with the value to optimize performance.

2. Why does NWChem fail when I set n_vcpus=1?

If NWChem was compiled with the MPI-PR (MPI Progress Rank) backend for ARMCI, it requires at least two MPI ranks to run.
This is because one MPI rank is reserved as a progress rank (responsible for handling communication), and the remaining ranks are used for computation.

What this means for you

• If you set n_vcpus=1, NWChem will fail because there are no compute ranks left after reserving one for progress.
• The minimum is n_vcpus=2: one rank will serve as the progress rank, and the other as the compute rank.

Still can't find what you're looking for? Contact Us