search

PSims / BayesEoR

Code to estimate the power spectrum of redshifted 21-cm emission from interferometric observations, within a Bayesian forward modelling framework.
https://bayeseor.readthedocs.io/en/latest/
BSD 3-Clause "New" or "Revised" License
2 stars 1 forks source link

[joss-review] memory requirement of example run #22

Closed zonca closed 2 months ago

zonca commented 2 months ago

the first step of the analysis requires about 20 GB of RAM, it would be useful to specify it in the docs, I ran with 15 GB and the process was killed (probably for out of memory)

I'll also estimate memory requirement for the second step and add it here.

jburba commented 2 months ago

I've updated the "Test Dataset" docs under usage.rst with RAM requirements (b9af6b5). When I ran the test dataset on a slurm system, I see the matrix construction step required ~17 GB of RAM according to the MaxRSS column in the slurm database and the power spectrum analysis required ~12 GB of RAM. The RAM requirement for the matrix construction being ~17 GB is probably why your job was killed with 15 GB of RAM but succeeded with 20 GB.

zonca commented 2 months ago

it seems like the docs website is not automatically updating: https://bayeseor.readthedocs.io/en/latest/usage.html#running-the-power-spectrum-analysis

as you estimated the power spectrum requirement for power spectrum analysis, I think it would be a good idea to add that number in. Do you also know how much GPU memory is needed?

jburba commented 2 months ago

Ah, yes. I had to manually rebuild the docs (I haven't automated the docs building yet). The documentation should be updated now.

Regarding GPU memory, I believe the routing we are using in MAGMA is capable of partial i/o so it should be capable of running on a wide range of hardware. I can try re-running the test data analysis and using e.g. nvidia-smi to try and get some information on GPU memory usage if necessary.