search

skuschel / postpic

The open-source particle-in-cell post-processor.
GNU General Public License v3.0
67 stars 27 forks source link

pyfftw cache causes memory overflow #144

Closed skuschel closed 6 years ago

skuschel commented 6 years ago

Using the experimental.kspace_propagate_adaptive continuously increases memory usage while propagating over multiple steps up to more than 60GB of memory. It turned out, that adding

import pyfftw
pyfftw.interfaces.cache.disable()

before the propagation loop fixed the problem and the memory consumption was kept constant at around 5GB.

Ablinne commented 6 years ago

Maybe it is sufficient to reduce the cache's keepalive time. It is set in datahandling.py line ca. 100: fftw_cache.set_keepalive_time(3600). The default value was quite small, I just increased it because I guessed it may be a good idea. Maybe that guess was wrong...

skuschel commented 6 years ago

It seems that the interface cache of pyfftw caches the PyFFTW objects and along with them all the data transformed. It is designed to reduce the (python)-overhead when creating a new PyFFTW object:

https://hgomersall.github.io/pyFFTW/pyfftw/interfaces/interfaces.html#pyfftw.interfaces.cache.enable

To me this seems beneficial only, when multiple fft with identical grid numbers and data type are performed, which is a VERY specialized use case to save VERY little time on each transformation. Thats why the default time is set to 0.1sec.

It seems from the documentation, that the planner wisdom is cached anyways (which is the big timesaver). However, that may be a good idea to check first.