Running Distributed Quality-Diversity Algorithms on HPC
Bryon Tjanaka, ICAROS Lab | 16 February 2021
Article
Slides
Overview
- Distributed QD Algorithms
- Hardware
- Distributed Computations with Dask
- Additional Considerations
- Further Resources
- Demo
Distributed QD Algorithms
Regular QD
archive = init_archive()
for n in range(generations):
solutions = create_new_solutions(archive)
evaluate(solutions)
insert_into_archive(solutions)
Distributed QD
archive = init_archive()
for n in range(generations):
solutions = create_new_solutions(archive)
distribute_evaluations(solutions)
retrieve_evaluation_results()
insert_into_archive(solutions)
Hardware
Single-Machine
HPC
ssh USCNETID@discovery.usc.edu
1 node
20+ cores
100+ GB RAM
sbatch job.slurm
job.slurm
Slurm config
(Bash) commands
Distributing Computation with Dask
- Python Library
- Workers + Scheduler
- TCP Connections
Running an Experiment on Dask
- Start scheduler job
- Start worker jobs
- Connect workers to scheduler
- Connect experiment to scheduler
- Repeatedly request evaluations
Additional Considerations
Integrating Dask into an Experiment
Environments / Containers
Configuration
Cluster
Algorithm
Running Robust Experiments
def evaluation():
logging.info("...")
Logging and Monitoring
Reloading
save(reload_data, "tmp.pkl")
rename("tmp.pkl", "reload.pkl")
Further Resources
Demo