k-Wave

1. Anastasiia
   

   Member
   Joined: Nov '19
   Posts: 20

   Hi Bradley,
   I have just run a simulation with a 306×306×1320 space grid (dx=dy=dz: 8.3333e-05) and a t_array: [1×9463 double]. i simulate a circular pressure source emitting a 20-cycle tone burst at 4.5 MHz.
   After the message that the simulation has been completed(''simulation completed in 81hours 47min 35.9105s'') I get the following error:
   ''Error using ifft
   Out of memory. Type "help memory" for your options.

   Error in gaussianFilter (line 81)
   signal = real(ifft(ifftshift(...

   Error in kspaceFirstOrder3D (line 1345)
   sensor_data.p = gaussianFilter(sensor_data.p, 1/kgrid.dt, sensor.frequency_response(1),
   sensor.frequency_response(2));''
   And the strange part about this is that the video with the wave propagation has still been recorded, so it seems like the simulation has indeed been completed. Is there perhaps some way to retrieve this data? and what would be your advice on calculating if there is enough memory beforehand? I am somewhat far from computer simulations so my apologies for the stupid question.

   Posted 3 years ago #

2. Anastasiia
   

   Member
   Joined: Nov '19
   Posts: 20

   I ran another simulation, with a smallet time step, compared to the one described above(now i have 7089 time points) and i get the following ''Error in kspaceFirstOrder3D (line 976)
   - absorb_eta .* real(ifftn( absorb_nabla2 .* fftn(rhox + rhoy + rhoz) )) ...''.
   Again a memory issue.

   Posted 3 years ago #

3. vich
   

   Member
   Joined: Jul '19
   Posts: 19

   Hi Anastasiia,

   Regarding your first post: the simulation should indeed appear to be complete, as the main computational loop in kspaceFirstOrder3D.m, which iterates over time steps, completes before the error you have shown. The video file is also fully saved to disk (lines 1195, 1229) before the error is encountered. (I assume you are using v.1.3, based on your error messages.)

   I am not entirely sure how many (modified) copies of sensor_data.p are kept around in memory while the critical line 81 in gaussianFilter.m runs. I would suspect the original version (in kspaceFirstOrder3D() scope), plus one or two versions in gaussianFilter() scope (one to hold the temporary results of line 81, and perhaps another to be passed as the input to fft(), though I'm not sure whether this second copy is the same as kspaceFirstOrder3D()'s version or not), plus an indeterminate number of versions used for intermediate steps in the calculation (I am not sure how MATLAB optimizes all those chained commands). You may want to run an empirical test with a smaller simulation and monitor your RAM usage (with top(1), taskmgr.exe, etc.), in terms of multiples of sensor_data.p, to see how the program behaves over time. Afterwards, you'll need to reduce your simulation size so that even at peak RAM usage your machine can handle the necessary computations.

   Posted 3 years ago #

4. Bradley Treeby
   

   Developer
   Joined: Mar '10
   Posts: 1,643

   Hi Anastasiia,

   In addition to the suggestions from vich, you could try removing sensor.frequency_response and applying this yourself using gaussianFilter after the simulation has completed. That way a lot of memory will be freed. In the extreme case, you could clear everything except your sensor data. This would also allow you to switch to using the C++ code, which will make the simulations faster also.

   Brad

   Posted 3 years ago #

RSS feed for this topic

Reply

You must log in to post.