This is the support structure minimization filter. So I need to go layer-by-layer from the bottommost slice of the array and update it as I move up. Every slice needs the updated values below that slice.
Vysakh From: Blaise Bourdin <[email protected]> Sent: Tuesday, January 17, 2023 4:47 PM To: Venugopal, Vysakh (venugovh) <[email protected]> Cc: Barry Smith <[email protected]>; [email protected] Subject: Re: [petsc-users] about repeat of expensive functions using VecScatterCreateToAll External Email: Use Caution What type of filter are you implementing? Convolution filters are expensive to parallelize since you need an overlap of the size of the support of the filter, but it may still not be worst than doing it sequentially (typically the filter size is only one or 2 element diameters). Or you may be able to apply the filter in Fourier space. PDE-filters are typically elliptic and can be parallelized. Blaise On Jan 17, 2023, at 4:38 PM, Venugopal, Vysakh (venugovh) via petsc-users <[email protected]<mailto:[email protected]>> wrote: Thank you! I am doing a structural optimization filter that inherently cannot be parallelized. Vysakh From: Barry Smith <[email protected]<mailto:[email protected]>> Sent: Tuesday, January 17, 2023 3:28 PM To: Venugopal, Vysakh (venugovh) <[email protected]<mailto:[email protected]>> Cc: [email protected]<mailto:[email protected]> Subject: Re: [petsc-users] about repeat of expensive functions using VecScatterCreateToAll External Email: Use Caution On Jan 17, 2023, at 3:12 PM, Venugopal, Vysakh (venugovh) via petsc-users <[email protected]<mailto:[email protected]>> wrote: Hi, I am doing the following thing. Step 1. Create DM object and get global vector ‘V’ using DMGetGlobalVector. Step 2. Doing some parallel operations on V. Step 3. I am using VecScatterCreateToAll on V to create a sequential vector ‘V_SEQ’ using VecScatterBegin/End with SCATTER_FORWARD. Step 4. I am performing an expensive operation on V_SEQ and outputting the updated V_SEQ. Step 5. I am using VecScatterBegin/End with SCATTER_REVERSE (global and sequential flipped) to get V that is updated with new values from V_SEQ. Step 6. I continue using this new V on the rest of the parallelized program. Question: Suppose I have n MPI processes, is the expensive operation in Step 4 repeated n times? If yes, is there a workaround such that the operation in Step 4 is performed only once? I would like to follow the same structure as steps 1 to 6 with step 4 only performed once. Each MPI rank is doing the same operations on its copy of the sequential vector. Since they are running in parallel it probably does not matter much that each is doing the same computation. Step 5 does not require any MPI since only part of the sequential vector (which everyone has) is needed in the parallel vector. You could use VecScatterCreateToZero() but then step 3 would require less communication but step 5 would require communication to get parts of the solution from rank 0 to the other ranks. The time for step 4 would be roughly the same. You will likely only see a worthwhile improvement in performance if you can parallelize the computation in 4. What are you doing that is computational intense and requires all the data on a rank? Barry Thanks, Vysakh Venugopal --- Vysakh Venugopal Ph.D. Candidate Department of Mechanical Engineering University of Cincinnati, Cincinnati, OH 45221-0072 — Canada Research Chair in Mathematical and Computational Aspects of Solid Mechanics (Tier 1) Professor, Department of Mathematics & Statistics Hamilton Hall room 409A, McMaster University 1280 Main Street West, Hamilton, Ontario L8S 4K1, Canada https://www.math.mcmaster.ca/bourdin | +1 (905) 525 9140 ext. 27243
