http://www.k-wave.org/forum/topic/memory-requirments-for-larfe-grid Support for the k-Wave MATLAB toolbox en-US Tue, 12 May 2026 23:28:24 +0000 http://bbpress.org/?v=1.0.2 q http://www.k-wave.org/forum/search.php http://www.k-wave.org/forum/topic/memory-requirments-for-larfe-grid#post-5507 Tue, 17 May 2016 15:31:31 +0000 umitarabul 5507@http://www.k-wave.org/forum/ <p>Thanks for the answer Dr. Treeby. I will look at those two studies.</p> <p>Umit </p> http://www.k-wave.org/forum/topic/memory-requirments-for-larfe-grid#post-5503 Sun, 15 May 2016 15:50:22 +0000 Bradley Treeby 5503@http://www.k-wave.org/forum/ <p>Hi umitarabul,</p> <p>Anthony Grisey did something similar to this in <a href="http://jtultrasound.biomedcentral.com/articles/10.1186/s40349-016-0056-9">this paper</a>. There are a few related forum posts, for example, <a href="http://www.k-wave.org/forum/topic/simulation-layer-by-layer">here</a>.</p> <p>Brad. </p> http://www.k-wave.org/forum/topic/memory-requirments-for-larfe-grid#post-5496 Tue, 10 May 2016 13:13:39 +0000 umitarabul 5496@http://www.k-wave.org/forum/ <p>Hi,</p> <p>Is it possible to divide the volume in pieces and run simulations separately and superpose the solutions? Is there a way to define the output of one block as input to the neighbor block? If we neglect non-linearity for example? </p> http://www.k-wave.org/forum/topic/memory-requirments-for-larfe-grid#post-5445 Mon, 04 Apr 2016 21:05:05 +0000 Bradley Treeby 5445@http://www.k-wave.org/forum/ <p>Hi Benji,</p> <p>This equation only applies to the C++ code. There is a significant difference between the two codes, with the C++ code being heavily optimised. I would definitely use this for larger problems if at all possible. Also, if you're trying to conserve memory, it's a good idea to run the C++ code outside of MATLAB using the <code>SaveToDisk</code> option rather than the inbuilt <code>kspaceFirstOrder3DC</code> function. This avoids the variables being in memory twice (once in MATLAB and once in C++).</p> <p>Hope that helps,</p> <p>Brad. </p> http://www.k-wave.org/forum/topic/memory-requirments-for-larfe-grid#post-5433 Fri, 18 Mar 2016 14:26:47 +0000 benji 5433@http://www.k-wave.org/forum/ <p>hi,<br /> Bradly - thank u. is the equation relevant also to the MATLAB case?<br /> is their a difference in memory requirments between the matlab and C codes?<br /> I tried running a hetrogenious grid of 768X768X768 with the MATLAB code using 45GB of physical memory and 45GB of virtual memory (90gb in total)<br /> it crushes because of memory issues, while according to equation 4.1 it requires only 25-30GB. </p> <p>marlen - sorry, but i have no magical solution, and as it seems so does Brad:). got to customize your simulation to meet practical limitations... I for instance will probably have to compromise on grid size and work with lower frequency. </p> http://www.k-wave.org/forum/topic/memory-requirments-for-larfe-grid#post-5419 Wed, 16 Mar 2016 13:23:31 +0000 Bradley Treeby 5419@http://www.k-wave.org/forum/ <p>Hi benji,</p> <p>If you are using the C++ code, there is a formulae for the memory requirements given in Eq. 4.1 in the <a href="http://www.k-wave.org/documentation.php">user manual</a> (section 4.9). It neglects a few small matrices, but it should give you a reasonable estimate.</p> <p>Brad. </p> http://www.k-wave.org/forum/topic/memory-requirments-for-larfe-grid#post-5418 Wed, 16 Mar 2016 12:25:18 +0000 marlenep 5418@http://www.k-wave.org/forum/ <p>Hi Benji,<br /> I have your same problem...have you found any solution? thank you very much for any suggestion </p> http://www.k-wave.org/forum/topic/memory-requirments-for-larfe-grid#post-5415 Sun, 13 Mar 2016 19:34:27 +0000 benji 5415@http://www.k-wave.org/forum/ <p>hi,<br /> Jiri - although u havent made me any more optimistic, thank u very much for your reply.</p> <p>a 2D simulation won't be helpful in that case. It must be a 3D one.<br /> as i said, i have about 200GB RAM memory available (includes both physical and virtual memory) at max. i will have to do the best under those contrains.</p> <p>how can i calculte memory requirments as a function of grid size? is their some thumb rule or simlpe equation to evluate that? </p> http://www.k-wave.org/forum/topic/memory-requirments-for-larfe-grid#post-5414 Sat, 12 Mar 2016 22:34:26 +0000 Jiri Jaros 5414@http://www.k-wave.org/forum/ <p>Hi Benji,<br /> It may be possible to run such a simulation with the MPI code, however, it would be extremely expensive!</p> <p>To simulate the wave propagating over 3000^3, you would need about 3TB of RAM. The simulation would have to run over at least 50k time steps. Using 3000 compute cores, and predicting that a time step would last 20s, you would have to wait for 277 hours for the simulation to complete. Taking into account 1 core-hours is for $0.1, you’d have to pay $83,000, which is, you’ll agree, quite impractical. </p> <p>We’re working on the faster GPU code, we hope could get the simulation cost below $10k, but still it’s really expensive for routine runs.</p> <p>I’d recommend you to start with a 2D slice. That would be much cheaper and faster. </p> http://www.k-wave.org/forum/topic/memory-requirments-for-larfe-grid#post-5407 Sun, 28 Feb 2016 23:42:37 +0000 benji 5407@http://www.k-wave.org/forum/ <p>Hi again,<br /> as mentioned in my previous post - I am simulating acoustic propagation in a humen body phantom. the desired max frequancy is around f=3MHz, and the 3D dimensions should be at minimum 12cmX12cmX12cm.</p> <p>in order to simulate body surface returns, the computational grid includs air with a sound speed of around 300m/s.</p> <p>That makes the greed spacing: dx=dy=dz= Cmin/[(num. of greed points per wave length)*fmax]<br /> I saw on the k-wave guide it is recomended to have num. of greed points per wave length at least 3-4 in a heteroginious medium. Cmin is 300m/s as mentioned above.<br /> that makes dx=dy=dz=~0.03mm.<br /> that meens the grid size shouldvbe about 3000X3000X3000 big.</p> <p>What would the memory requirments be to handle such a grid?<br /> i saw somwhere on the forum here that a 256X256X256 would require 6GB, so linearly interpulating i guess I would need some 12,000GB... :-o<br /> 1. am i right?<br /> 2. any practical ideas how to run such a simulation given a powerful university cluster with up to some 200GB or so ?</p> <p>thank u,<br /> Benji </p>