Automation scripts for the flatsurf stack to survey large sets of objects.

To perform a full survey, use flatsurvey. To investigate a single object, run flatsurvey-worker.

To play with this tool locally, you can run pixi shell to make sure that you have all the required dependencies installed in the correct versions.

Here, we check that the (1, 1, 1, 6) quadrilateral has dense orbit closure::

flatsurvey-worker ngon -a 1 -a 1 -a 1 -a 6 orbit-closure --deform

Here is a typical survey that collects data about triangles, quadrilaterals, and pentagons:

mkdir -p surveyname
nice flatsurvey ngons --vertices 3 ngons --vertices 4 ngons --vertices 5 orbit-closure --deform json --prefix=./surveyname

This will run indefinitely, so you want to stop this at some point with C-c.

To use the results of this survey to speed up a future survey, we collect the results in one survey-results/orbit_closure.json:

flatsurvey-maintenance join --outdir survey-results surveyname/*.json

And then use that JSON file as a database for the next run:

nice flatsurvey ngons --vertices 3 ngons --vertices 4 ngons --vertices 5 orbit-closure --deform local-cache --json orbit-closure.json json --prefix=./surveyname

We provide example scripts that run surveys on the PlaFRIM cluster. It should be easy to adapt these to work on any cluster that can run a dask workload.

Setup this pixi repository for the scheduler and the workers:

sh plafrim/create-env.sh

Spawn a dask scheduler (here we call our survey flatsurvey-3):

sh plafrim/spawn-scheduler.sh flatsurvey-3

Reserve resources in your cluster:

salloc --ntasks=768 --time=13:30:00 --constraint="diablo|bora|brise|sirocco|zonda|miriel|souris|kona"

Spawn the workers:

for host in $(scontrol show hostnames); do sleep 1; srun --nodes=1 --ntasks=1 --exclusive -w $host sh plafrim/provision-env.sh flatsurvey-3 & done
wait
srun sh plafrim/spawn-worker.sh flatsurvey-3

Start the survey:

sh plafrim/survey.sh flatsurvey-3 "ngons --vertices 3"

Post-process the survey:

sh plafrim/postprocess.sh flatsurvey-3

...

• For large surveys, RAM might become the limiting factor. It appears that we are not actually leaking memory but are hit by memory fragmentation during the Boshernitzan criterion. The issue can be fully mitigated by replacing malloc with Mesh, i.e., setting LD_PRELOAD=/path/to/libmesh.so.
• SageMath (or one of its dependencies) might decide that it's beneficial to parallelize things further. However, this can easily overload a system. Typically, some linear algebra might spawn a process for each CPU on the system which could then easily lead to CPU² many processes if you run a lengthy survey. To disable this behaviour, set MKL_NUM_THREADS=1 OMP_NUM_THREADS=1 SAGE_NUM_THREADS=1.