`cache` issues
Unknown behaviors and bugs of new cache feature
-
cachework flows before sending to MPI ranks - Size of saved
cachegrowing unbounded - cache requiring recompilation for altered meshes, census time, and census bank buffer
After running in Numba mode and getting the compilation cached, we sometimes want to run in pure Python mode again (e.g., to do Python profiling, to compare Python and Numba error messages, etc). Currently, it seems that cached compilation will be used even if we run in Python mode. Is it possible to bypass or ignore the cached compilation in pure Python mode? @jpmorgan98
This is very odd behavior. Like even if it's in Python mode it will run as if in Numba? My only thought is to delete the cache folder all together. It should be in /__pycache__ in your execution directory
It is found that, since we only put the cache decorator in the loop.py, if we make changes in kernel.py, the cache will not recompile.
Has this been resolved, @braxtoncuneo @jpmorgan98 ?
Do we need to document the new caching mechanism?
Also, is it possible to add a regression test to check if the new caching options (--caching, etc.) work?
Do we have the documentation on how to use the new caching options?