<div dir="ltr"><div dir="ltr"><br></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Fri, Sep 3, 2021 at 1:57 AM Viktor Nazdrachev <<a href="mailto:numbersixvs@gmail.com">numbersixvs@gmail.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black">Hello, Lawrence!<br>
Thank you for your response!</span></p>
<p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black">I attached log files (txt files with convergence
behavior and RAM usage log in separate txt files) and resulting table with
convergence investigation data(xls). Data for main non-regular grid with 500K cells
and heterogeneous properties are in 500K folder, whereas data for simple
uniform 125K cells grid with constant properties are in 125K folder. </span></p><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black"><br></span></p><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>><i> On 1 Sep 2021, at 09:42, </i></span><i><span style="color:black">Наздрачёв</span></i><i><span style="color:black"> </span><span style="color:black">Виктор</span></i><i><span lang="EN-US" style="color:black"> <</span><span style="color:black"><a href="https://lists.mcs.anl.gov/mailman/listinfo/petsc-users" style="color:blue" target="_blank"><span lang="EN-US" style="color:black">numbersixvs at gmail.com</span></a></span></i><i><span lang="EN-US" style="color:black">> wrote:</span></i></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>><i> </i></span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>><i> I have a 3D elasticity problem with heterogeneous properties.</i></span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">> </span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>What does your coefficient variation look like? How large is the contrast?</span></pre><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black">Young modulus varies from 1 to 10 GPa, Poisson ratio varies
from 0.3 to 0.44 and density – from 1700 to 2600 kg/m^3.</span></p></div></blockquote><div><br></div><div>That is not too bad. Poorly shaped elements are the next thing to worry about. Try to keep the aspect ratio below 10 if possible.</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black"> </span></p><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>><i> There is unstructured grid with aspect ratio varied from 4 to 25. Zero Dirichlet BCs are imposed on bottom face of mesh. Also, Neumann (traction) BCs are imposed on side faces. Gravity load is also accounted for. The grid I use consists of 500k cells (which is approximately 1.6M of DOFs).</i></span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>><i> </i></span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>><i> The best performance and memory usage for single MPI process was obtained with HPDDM(BFBCG) solver and bjacobian + ICC (1) in subdomains as preconditioner, it took 1 m 45 s and RAM 5.0 GB. Parallel computation with 4 MPI processes took 2 m 46 s when using 5.6 GB of RAM. This because of number of iterations required to achieve the same tolerance is significantly increased.</i></span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">> </span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>How many iterations do you have in serial (and then in parallel)?</span></pre><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">Serial run is required 112 iterations to reach convergence (log_hpddm(bfbcg)_bjacobian_icc_1_mpi.txt), parallel run with 4 MPI – 680 iterations.</span></pre><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black">I attached log files for all simulations (txt files
with convergence behavior and RAM usage log in separate txt files) and
resulting table with convergence/memory usage data(xls). Data for main
non-regular grid with 500K cells and heterogeneous properties are in 500K
folder, whereas data for simple uniform 125K cells grid with constant
properties are in 125K folder. </span></p><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>><i> I`ve also tried PCGAMG (agg) preconditioner with IC</i></span><i><span style="color:black">С</span></i><i><span lang="EN-US" style="color:black"> (1) sub-precondtioner. For single MPI process, the calculation took 10 min and 3.4 GB of RAM. To improve the convergence rate, the nullspace was attached using MatNullSpaceCreateRigidBody and MatSetNearNullSpace subroutines. This has reduced calculation time to 3 m 58 s when using 4.3 GB of RAM. Also, there is peak memory usage with 14.1 GB, which appears just before the start of the iterations. Parallel computation with 4 MPI processes took 2 m 53 s when using 8.4 GB of RAM. In that case the peak memory usage is about 22 GB.</span></i></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">> </span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>Does the number of iterates increase in parallel? Again, how many iterations do you have?</span></pre><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><pre style="text-align:justify;margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">For case with 4 MPI processes and attached nullspace it is required 177 iterations to reach convergence (you may see detailed log in log_hpddm(bfbcg)_gamg_nearnullspace_4_mpi.txt). For comparison, 90 iterations are required for sequential run(log_hpddm(bfbcg)_gamg_nearnullspace_1_mpi.txt).</span></pre></div></blockquote><div><br></div><div>Again, do not use ICC. I am surprised to see such a large jump in iteration count, but get ICC off the table.</div><div><br></div><div>You will see variability in the iteration count with processor count with GAMG. As much as 10% +-. Maybe more (random) variability , but usually less.</div><div><br></div><div>You can decrease the memory a little, and the setup time a lot, by aggressively coarsening, at the expense of higher iteration counts. It's a balancing act.</div><div><br></div><div>You can run with the defaults, add '-info', grep on GAMG and send the ~30 lines of output if you want advice on parameters.</div><div><br></div><div>Thanks,</div><div>Mark</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>><i> Are there ways to avoid decreasing of the convergence rate for bjacobi precondtioner in parallel mode? Does it make sense to use hierarchical or nested krylov methods with a local gmres solver (sub_pc_type gmres) and some sub-precondtioner (for example, sub_pc_type bjacobi)?</i></span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">> </span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>bjacobi is only a one-level method, so you would not expect process-independent convergence rate for this kind of problem. If the coefficient variation is not too extreme, then I would expect GAMG (or some other smoothed aggregation package, perhaps -pc_type ml (you need --download-ml)) would work well with some tuning.</span></pre><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black">Thanks for idea, but, unfortunately, ML cannot be
compiled with 64bit integers (It is extremely necessary to perform computation
on mesh with more than 10M cells).</span></p><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>If you have extremely high contrast coefficients you might need something with stronger coarse grids. If you can assemble so-called Neumann matrices (</span><span style="color:black"><a href="https://petsc.org/release/docs/manualpages/Mat/MATIS.html#MATIS" style="color:blue" target="_blank"><span lang="EN-US" style="color:black">https://petsc.org/release/docs/manualpages/Mat/MATIS.html#MATIS</span></a></span><span lang="EN-US" style="color:black">) then you could try the geneo scheme offered by PCHPDDM.</span></pre><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black"> </span></p><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black">I found strange convergence behavior for HPDDM
preconditioner. For 1 MPI process BFBCG solver did not converged
(log_hpddm(bfbcg)_pchpddm_1_mpi.txt), while for 4 MPI processes computation was
successful (1018 to reach convergence, log_hpddm(bfbcg)_pchpddm_4_mpi.txt).</span></p><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black">But it should be mentioned that stiffness matrix was
created in AIJ format (our default matrix format in program).</span></p><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black">Matrix conversion to MATIS format via MatConvert </span><span lang="EN-US" style="font-family:"Courier New";color:black">subroutine resulted in losing of convergence for both
serial and parallel run.</span><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black"></span></p><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-family:"Courier New";color:black"><br></span></p><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>><i> Is this peak memory usage expected for gamg preconditioner? is there any way to reduce it?</i></span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">> </span></pre><pre style="margin:0cm 0cm 0.0001pt 35.4pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black">>I think that peak memory usage comes from building the coarse grids. Can you run with `-info` and grep for GAMG, this will provide some output that more expert GAMG users can interpret.</span> </pre><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span></pre><pre style="margin:0cm 0cm 0.0001pt;font-size:10pt;font-family:"Courier New""><span lang="EN-US" style="color:black"> </span>Thanks, I`ll try to use a strong threshold only for coarse grids.</pre>
<p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black"> </span></p>
<p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black">Kind regards,</span></p>
<p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black"> </span></p>
<p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black">Viktor Nazdrachev</span></p>
<p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black"> </span></p>
<p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black">R&D senior researcher</span></p>
<p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black"> </span></p>
<p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black">Geosteering Technologies LLC</span></p><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black"> </span></p><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black"> </span></p><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black"> </span></p><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black">
</span></p><p class="MsoNormal" style="margin:0cm;line-height:normal;font-size:11pt;font-family:Calibri,sans-serif"><span lang="EN-US" style="font-size:10pt;font-family:"Courier New";color:black"> </span></p></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">ср, 1 сент. 2021 г. в 12:02, Lawrence Mitchell <<a href="mailto:wence@gmx.li" target="_blank">wence@gmx.li</a>>:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><br>
<br>
> On 1 Sep 2021, at 09:42, Наздрачёв Виктор <<a href="mailto:numbersixvs@gmail.com" target="_blank">numbersixvs@gmail.com</a>> wrote:<br>
> <br>
> I have a 3D elasticity problem with heterogeneous properties.<br>
<br>
What does your coefficient variation look like? How large is the contrast?<br>
<br>
> There is unstructured grid with aspect ratio varied from 4 to 25. Zero Dirichlet BCs are imposed on bottom face of mesh. Also, Neumann (traction) BCs are imposed on side faces. Gravity load is also accounted for. The grid I use consists of 500k cells (which is approximately 1.6M of DOFs).<br>
> <br>
> The best performance and memory usage for single MPI process was obtained with HPDDM(BFBCG) solver and bjacobian + ICC (1) in subdomains as preconditioner, it took 1 m 45 s and RAM 5.0 GB. Parallel computation with 4 MPI processes took 2 m 46 s when using 5.6 GB of RAM. This because of number of iterations required to achieve the same tolerance is significantly increased.<br>
<br>
How many iterations do you have in serial (and then in parallel)?<br>
<br>
> I`ve also tried PCGAMG (agg) preconditioner with ICС (1) sub-precondtioner. For single MPI process, the calculation took 10 min and 3.4 GB of RAM. To improve the convergence rate, the nullspace was attached using MatNullSpaceCreateRigidBody and MatSetNearNullSpace subroutines. This has reduced calculation time to 3 m 58 s when using 4.3 GB of RAM. Also, there is peak memory usage with 14.1 GB, which appears just before the start of the iterations. Parallel computation with 4 MPI processes took 2 m 53 s when using 8.4 GB of RAM. In that case the peak memory usage is about 22 GB.<br>
<br>
Does the number of iterates increase in parallel? Again, how many iterations do you have?<br>
<br>
> Are there ways to avoid decreasing of the convergence rate for bjacobi precondtioner in parallel mode? Does it make sense to use hierarchical or nested krylov methods with a local gmres solver (sub_pc_type gmres) and some sub-precondtioner (for example, sub_pc_type bjacobi)?<br>
<br>
bjacobi is only a one-level method, so you would not expect process-independent convergence rate for this kind of problem. If the coefficient variation is not too extreme, then I would expect GAMG (or some other smoothed aggregation package, perhaps -pc_type ml (you need --download-ml)) would work well with some tuning.<br>
<br>
If you have extremely high contrast coefficients you might need something with stronger coarse grids. If you can assemble so-called Neumann matrices (<a href="https://petsc.org/release/docs/manualpages/Mat/MATIS.html#MATIS" rel="noreferrer" target="_blank">https://petsc.org/release/docs/manualpages/Mat/MATIS.html#MATIS</a>) then you could try the geneo scheme offered by PCHPDDM.<br>
<br>
> Is this peak memory usage expected for gamg preconditioner? is there any way to reduce it?<br>
<br>
I think that peak memory usage comes from building the coarse grids. Can you run with `-info` and grep for GAMG, this will provide some output that more expert GAMG users can interpret.<br>
<br>
Lawrence<br>
<br>
</blockquote></div>
</blockquote></div></div>