Skip to content

Evolutionary Searches

Danger

This app is still under early testing and not really for use outside of the Warren Lab

Danger

⚠⚠ Make sure you are using a cloud database if your workers are on separate machines or an HPC cluster. ⚠⚠

The default database backend (sqlite) is not built for parallel connections from different computers. Your calculations will be slower and error-prone with sqlite.

If you are seeing the error database is locked, then you have exceeded the capabilities of sqlite.

Warning

Just like with earlier tutorials, a working VASP installation is required.

The quick tutorial

1) View input options and default settings of the structure-prediction.toolkit.fixed-composition workflow 

simmate workflows explore

2) (optional) If you wish for prototype structures or known materials to be included at the start of your search, make sure you have loaded the data into your cloud database. This was covered in an earlier tutorial.

3) Build our input yaml file (e.g. my_search.yaml). Be sure to look at the full example for best-practices.

workflow_name: structure-prediction.toolkit.fixed-composition
composition: Ca8N4F8

workflow_name: structure-prediction.toolkit.fixed-composition
composition: Ca8N4F8

subworkflow_kwargs:  # (1)
    command: mpirun -n 8 vasp_std > vasp.out  # (2)
    compress_output: true  # (3)
    # see 'simmate workflows explore' on `relaxation.vasp.staged`
    # for other optional subworkflow_kwargs

sleep_step: 300  # (4)
nsteadystate: 100  # (5)

# see 'simmate workflows explore' output for other optional inputs
  1. One of the default input parameters is subworkflow: relaxation.vasp.staged, which is used to analyze each structure. We can modify how this workflow behaves with the subworkflow_kwargs setting.
  2. The command that each relaxation step will be called with.
  3. You likely won't be reading through the output files of each calculation, and these resuklts will take up a lot of file space. This setting will convert completed calculations to zip files and help save on disk space.
  4. The status of individual structures & the writing of output files is checked on a cycle. If you know your average relaxation will take 30 minutes, then you likely don't need to check/update every 60 seconds (the default). Longer sleep steps help reduce database load.
  5. By default, nsteadystate is set to 40, which means the search will maintain 40 total calculations in the queue at all times -- regardless of the number of workers available. The number of workers (not nsteadystate) controls the number of parallel calculations. This value should only be changed if you want more than 40 calculations ran in parallel. This means >40 workers must be running. Therefore, increase this value if you desire but do NOT decrease nsteadystate.

4) Submit the workflow settings file, which will start scheduling jobs. 

simmate workflows run my_search.yaml

5) Start workers by submitting the start-worker command to a cluster's queue (e.g. SLURM) or whereever you'd like jobs to run. Submit as many workers as workflows that you want ran in parallel:

simmate engine start-worker

# (1)
simmate engine start-worker --close-on-empty-queue --nitems-max 10
  1. --close-on-empty-queue shuts down the worker when the queue is empty. This helps release computational resources for others. --nitems-max limits the number of workflows each worker will run. Short-lived workers help maintain the health of a cluster and also allow other SLURM jobs to cycle through the queue -- avoiding the hogging of resources.

6) Monitor the output and log files for any issues. Important error information can also be accessed in the command line: 

simmate engine stats
simmate engine error-summary

7) Submit new workers or cancel stale workers as needed.

  1. Follow the steps from the fixed-composition search above. The only difference here will be (i) the input yaml and (ii) follow-up searches. Make sure you read and understand best practices as well.

  2. Build your input yaml file: 

    workflow_name: structure-prediction.toolkit.chemical-system
chemical_system: Ca-N

  3. Submit the search 

    simmate workflows run my_search.yaml

  4. Submit workers as-needed 

    simmate engine start-worker