Finite Element Analysis with High Throughput Computing

Kevin Fang, Biomechatronics @ MIT Media Lab, March 2018

Distribute finite element analysis jobs to a cluster of computers running Linux using HTCondor. This specific repository applies to using FEBio. However, this idea can be applied to any project that has lots of files that need to be processed.

---

About

This repository uses HTCondor, high throughput compute software created by UW-Madison. The goal of this project is to create a network that distributes many complex software jobs (e.g. thousands of optimization problems) to lots of computers. This network does not do well with running a single, very intense workload (e.g. one 15 hour calculation).

---

Installation

INSTALL.md contains important installation instructions on how to set up the Condor network.
Complete list of steps to follow from INSTALL.md:

1. Set up a central manager (either using the Docker or native method). In condor_config.local, set DAEMON_LIST to MASTER, COLLECTOR, NEGOTIATOR, SCHEDD and optionally add STARTD if you want your central manager to be able to execute jobs to. Record the IP address of this computer. Condor dislikes hostnames, so you'll probably want to give this machine a static IP, as all the nodes will connect to it.
2. Set up execution machines (either Docker/native) using the instructions in INSTALL.md. Set slot definitions if needed. In condor_config.local, set the CONDOR_HOST variable to the IP address of the central manager. If you are deploying to a large configuration, it would be easiest to distribute the same configuration file to all the machines.
3. Once the native/Docker installations are all up, run condor_status on the central manager and make sure that you see all the machines.
4. Follow the Python script instructions below for creating jobs.

---

Using the Python scripts with Condor

generate_from_directory.py generates a single submission file that creates a job to run analysis on several .feb files in a directory. It takes the following arguments:

• --folder (required) points to the directory containing the .feb files (e.g. --folder ./files).
• --project_name (optional) sets a project name for the output job file. Default is job.sub.
• --ram_req (optional) sets a RAM requirement in megabytes for computers able to execute the job (e.g. --ram_req 1024).
• --cpu_req (optional) sets a CPU requirement in number of cores for computers able to execute the job (e.g. --cpu_req 4).
• --run - if this is included, immediately after generating the job submission file the program will submit the job to the cluster.

The generated files will be located in a directory called output/.

The script creates a .sub file to be submitted, and a file called febio.sh that contains the command for the execution machines to run.

For example, if you have .feb files located in a directory called feb_directory/, run python3 generate_from_directory.py --folder feb_directory --job sample_job and follow the instructions to generate a job submission. Once the script is run, feb_directory/ should look like this:

feb_directory/
├───model1.feb
├───model2.feb
├───model3.feb
└───output/
   ├───sample_job.sub
   └───febio.sh

• If you want to submit the job immediately, you would run python3 generate_from_directory.py --folder feb_directory --job sample_job --run instead.
• If you did not include --run, navigate to output/ and run condor_submit <submission file>.sub and wait for the cluster to run the analysis (the output of condor_submit should be <num_jobs> job(s) submitted to cluster <cluster_num>).
• Check on the status of the jobs with condor_q, or for more detailed analysis, write condor_q -analyze <cluster_num>. The output of the jobs can be found in a new directory called output/.
• After the analysis is finished, project_directory/ would look like this:

feb_directory/
├───model1.feb
├───model2.feb
├───model3.feb
└───output/
   ├───sample_job.sub
   ├───model1.txt, model1.log, model1.err + model1 outputs
   ├───model2.txt, model2.log, model2.err + model2 outputs
   ├───model3.txt, model3.log, model3.err + model3 outputs
   └───febio.sh

Expanding beyond FEBio

This repo can easily be modified to work with other command line processing software besides FEBio, but it involves knowing Docker, Python, and Condor, and requires a basic understanding of generate_from_directory.py. To use it with other software:

• Change the Dockerfile such that it installs the needed software in the image (e.g. if you're running MatLab, modify the Dockerfile so that it installs MatLab).
• Modify generate_from_directory.py so that script_file contains a command line argument to run the command line software. To feed in parameters, use $1.
• Modify the generate_job function in generate_from_directory.py so that it has the correct input/output names and redirects output correcty.

To do:

• Make a user friendly method of submitting jobs - perhaps a web server with Node.js/React.js?
• Create documentation for submitting jobs from other machines if the central manager is not easily accessible.