k-Wave

1. Hammerben1
   

   Member
   Joined: Jan '18
   Posts: 15

   Hi k-wave team:
   I am trying to repeat what you did in Treeby 2020, Experimental validation of k-Wave: Nonlinear wave propagation in layered, absorbing fluid media.

   How did you define the sensor_mask to run it sequentially for 8 subdomains?

   When I tried to define it as [1,1,Nz-8, Nx,Ny,Nz] and ask to record sensor_data.p, to be used as the source for the next subdomain, the data size was about 220GB, which is much higher than the RAM available to me.

   Also it seems that I can't record p_max with a separate sensor_mask. How did you record the sensor_data.p_max along with sensor_data.p with that large window (40x40x70mm)? The data size for sensor_data.p with that window is easily > 10TB

   These led me to think I am doing this wrong.

   Help greatly appreciated
   Thanks

   P.S. I have the NVIDIA V100 with 32GB and about 128GB RAM available. I have no issue running the first segment but it can't handle it when the data starts to transfer to disk

   Posted 3 years ago #

2. Hammerben1
   

   Member
   Joined: Jan '18
   Posts: 15

   An update on this thread:

   I successfully ran the simulation at a lower spatial resolution.
   I used a uniform 4-cycle burst for all source element.
   The results don't seem to match with the paper even for the free-field case, which uses a hydrophone measured back-projected plane as source plane.

   Does it mean we shouldn't define the source waveform as a single waveform in all of our simulations?

   Thanks

   Posted 3 years ago #

3. Elly Martin
   

   Developer
   Joined: Nov '14
   Posts: 4

   Hi Hammerben,

   In reply to your first post - the simulations for this paper were run using a multi-GPU version of k-Wave which splits the simulation domain across a number of GPUs, rather than as sequential simulations. The simulations run simultaneously and for all time steps with data exchanged between the subdomains. You can read more about how this works in J. Jaros, F. Vaverka, and B. E. Treeby, Supercomput. Frontiers Innov., vol. 3, no. 3, pp. 40–55, 2016. Unfortunately this version is not available yet but we aim to include it in a future release.

   Re the sensor data: setting the p_max_all flag records p_max over the whole domain. sensor_data.p is only recorded at the positions defined by the sensor mask. In the paper, this was a much smaller volume, for example, for the water simulations, the pressure time signals were recorded over a volume of 11x11x1701 grid points.

   In reply to you 2nd post: in the paper the source was defined using a back projected hydrophone measurement, resulting in non-uniform source waveform amplitudes across the source. It is likely that the shape of the measured waveform will also be different to the shape of the waveform generated by setting e.g a windowed sinusoidal burst. You can find more information on the differences which arise from uniform vs measured sources in E. Martin, Y. T. Ling, B. E. Treeby IEEE Trans. Ultrason. Ferroelectr. Freq. Control, 63 (10), pp. 1535–1542, 2016.

   If you wish to compare using the measured data (including the source characterization), it is available here:
   https://rdr.ucl.ac.uk/articles/dataset/Experimental_Validation_of_k-Wave_Nonlinear_Wave_Propagation_in_Layered_Absorbing_Fluid_Media/14216717

   Elly

   Posted 3 years ago #

RSS feed for this topic

Reply

You must log in to post.