ch-sim

A framework for running simulations with ADCIRC on HPC resources.

Description

This repository contains a collection of tools to easily benchmark, prototype and run ensemble simulations with ADCIRC. One purpose of the library is to automate common workflows and tasks involved in simulations - such as benchmark runs, parameter scans, editing parameter files, hostarts, etc. The goal is to provide a framework for running simulations that supports complex workflows such as MCMC, Metropolis Sampling, Bayesian inversion, etc.

Setup on TACC

While the framework should work with any slurm-type system, it is currently targeted for use on TACC's resources. Below is a guide to get started on TACC.

1. Download Miniconda for TACC wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
2. Run the installer bash Miniconda3-latest-Linux-x86_64.sh. When prompted for the installation directory, be sure to specify something not in the home partition, e.g. ($WORK/minconda3). This will prevent conda packages from taking up space in the limited home partition. When asked if you want to initialize conda, respond 'yes'.
3. Refresh the shell by running source .bashrc and create a new Conda environment by running conda create ch-sim.
4. Activate the new environment conda activate ch-sim and clone the ch-sim repo git clone https://github.com/UT-CHG/ch-sim.git.
5. Install pip conda install pip.
6. Install ch-sim cd ch-chim && pip install .

Example Usage

The simplest workfow is running a single ADCIRC simulation. Suppose we create a file named test.py below:

from ch_sim import BaseSimulator
import os

# make sure to replace with your allocation here
allocation = "ADCIRC"
# get path to WORK directory
workdir = os.path.expandvars("$WORK")

if __name__ == "__main__":
  sim = BaseSimulator(allocation=allocation)

  sim.run(
      inputs_dir=workdir+"/inputs", # directory with ADCIRC input files (fort.14, ....)
      outputs_dir=workdir+"/outputs", # directory to copy outputs
      execs_dir=workdir+"/execs", # directory with padcirc and adcprep executables
      node_count=10, # number of nodes to run the job on
      processors_per_node=56, # set to the number of CPUs per node on the target system (Frontera here)
      runtime=4, # number of hours
      queue="normal"
  )

Submitting the simulation is as simple as running the above Python file. python3 test.py. When the simulation completes, outputs will be copied to the directory specified via the argument outputs_dir.

The second most common workflow is a parameter sweep. In this workflow, a potentially large number of ADCIRC simulations need to be run with varying inputs. Frequently, these simulations share common input files (e.g. the fort.14 mesh file, etc.). We can easily submit a batch of ADCIRC simulations with varying inputs using the EnsembleSimulator class:

from ch_sim import EnsembleSimulator
import os

# make sure to replace with your allocation here
allocation = "ADCIRC"
# get path to WORK directory
workdir = os.path.expandvars("$WORK")

if __name__ == "__main__":
  sim = EnsembleSimulator(allocation=allocation)

  # Define run-specific input and output directories
  # Note that we don't need to have a full set of ADCIRC parameters in each input
  runs = [
     {
       "inputs_dir": workdir+"/harvey/inputs",
       "outputs_dir": workdir+"/harvey/outputs"
     },
     {
       "inputs_dir": workdir+"/katrina/inputs",
       "outputs_dir": workdir+"/katrina/outputs"
     },
     {
       "inputs_dir": workdir+"/ike/inputs",
       "outputs_dir": workdir+"/ike/outputs"
     }
  ]


  sim.run(
      # directory with ADCIRC input files that are shared between runs
      # If the same file is specified in both the base input directory and a run input directory,
      # the run file takes precedence
      inputs_dir=workdir+"/inputs",
      execs_dir=workdir+"/execs", # directory with padcirc and adcprep executables
      node_count=10, # number of nodes PER INDIVIDUAL RUN
      processors_per_node=56, # set to the number of CPUs per node on the target system (Frontera here)
      runtime=4, # expected runtime for a single ADCIRC simulations
      queue="normal",
      maxJobNodes=20, # max allowed nodes per job
      maxJobRuntime=4, # max allowed runtime per job
      # if all ADCIRC runs can't complete in a single job, multiple jobs will be submitted
      # tweaking maxJobNodes and maxJobRuntime can help optimize queue wait times and throughput
      runs=runs
  )

Running this file will submit two jobs - one that runs on 20 nodes for 4 hours and completes the first two ADCIRC runs, and one that handles just a single ADICRC run. If we want all simulations in the same job, we could increase maxJobNodes to 30, or alternatively increase maxJobRuntime.

Much more complex workflows are possible with the framework, including data assimilation, on-the-fly customization of the ADCIRC input files, and arbitrary pre/post processing steps. These are accomplished by subclassing the EnsembleSimulator and BaseSimulator classes. Future documentation is forthcoming. . .

Contact

Please reach out to

Benjamin Pachev <[email protected]> Carlos del-Castillo-Negrete <[email protected]>

to report bugs or suggest features.