Open kerim371 opened 6 days ago
It is fairly optimized yes but it will take longer by a bit. There is some option in pyrevolve
to control the memory but it's a bit of a pain to expose it to the user
It seems I can change n_checkpoints directly in J_adjoint in source code.
As I understand it should work faster if n_checkpoints
is big enough. Also if n_checkpoints == nt_comp
then the perfomance should be almost equal when JUDIOptions::optimal_checkpointing = false
right?
Hi,
I'm trying to use
optimal_checkpointing
to save memory. As I understand this option is aimed to recompute wavefields instead of keeping them in RAM or on the hardware.It seems that when I use this option my RAM memory is loaded only for 5-10% and the computations run about 10 times longer. Is it really optimized? Is there a way to control how much checkpoints is used or huw much memory is used during computations?