<div dir="ltr"><div>Thanks Jose for the help, also for me works now, and thank you Junchao for doing the test. I'll also think more from where this imbalance may come from, because there shouldn't be any, as far as I know.<br></div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">El lun., 1 jul. 2019 a las 18:36, Zhang, Junchao (<<a href="mailto:jczhang@mcs.anl.gov" target="_blank">jczhang@mcs.anl.gov</a>>) escribió:<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>
<div dir="ltr">
<div>Jose & Ale,</div>
<div> -ds_method 2 fixed the problem. I used PETSc master (f1480a5c) and slepc master(675b89d7) through --download-slepc. I used MKL /opt/intel/compilers_and_libraries_2018.1.163/linux/mkl/</div>
<div> I got the following results with 2048 processors. MatAssemblyEnd looks expensive to me. I am looking into it.</div>
<div><br>
</div>
<div><font face="courier new, monospace">--- Event Stage 5: Offdiag<br>
<br>
BuildTwoSidedF 1 1.0 1.5201e+007345.1 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 1 0 0 0 0 13 0 0 0 0 0<br>
MatAssemblyBegin 1 1.0 1.5201e+005371.4 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 1 0 0 0 0 13 0 0 0 0 0<br>
MatAssemblyEnd 1 1.0 1.7720e+00 1.0 0.00e+00 0.0 5.7e+04 1.3e+05 8.0e+00 2 0 0 0 1 39 0100100100 0<br>
VecSet 1 1.0 2.4695e-02 7.9 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0</font><br>
</div>
<div><br>
</div>
<div>--Junchao Zhang<br>
</div>
<br>
</div>
<br>
<div class="gmail_quote">
<div dir="ltr" class="gmail_attr">On Mon, Jul 1, 2019 at 4:28 AM Jose E. Roman <<a href="mailto:jroman@dsic.upv.es" target="_blank">jroman@dsic.upv.es</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">
You can try the following:<br>
- Try with a different DS method: -ds_method 1 or -ds_method 2 (see DSSetMethod)<br>
- Run with -ds_parallel synchronized (see DSSetParallel)<br>
If it does not help, send a reproducible code to slepc-maint<br>
<br>
Jose<br>
<br>
<br>
> El 1 jul 2019, a las 11:10, Ale Foggia via petsc-users <<a href="mailto:petsc-users@mcs.anl.gov" target="_blank">petsc-users@mcs.anl.gov</a>> escribió:<br>
> <br>
> Oh, I also got the same error when I switched to the newest version of SLEPc (using OpenBlas), and I don't know where it is coming from.<br>
> Can you tell me which version of SLEPc and PETSc are you using? And, are you using MKL?<br>
> Thanks for trying :)<br>
> <br>
> El vie., 28 jun. 2019 a las 16:57, Zhang, Junchao (<<a href="mailto:jczhang@mcs.anl.gov" target="_blank">jczhang@mcs.anl.gov</a>>) escribió:<br>
> Ran with 64 nodes and 32 ranks/node, met slepc errors and did not know how to proceed :(<br>
> <br>
> [363]PETSC ERROR: --------------------- Error Message --------------------------------------------------------------<br>
> [363]PETSC ERROR: Error in external library<br>
> [363]PETSC ERROR: Error in LAPACK subroutine steqr: info=0<br>
> [363]PETSC ERROR: See <a href="https://www.mcs.anl.gov/petsc/documentation/faq.html" rel="noreferrer" target="_blank">
https://www.mcs.anl.gov/petsc/documentation/faq.html</a> for trouble shooting.<br>
> [363]PETSC ERROR: Petsc Development GIT revision: v3.11.2-1052-gf1480a5c GIT Date: 2019-06-22 21:39:54 +0000<br>
> [363]PETSC ERROR: /tmp/main.x on a arch-cray-xc40-knl-opt named nid03387 by jczhang Fri Jun 28 07:26:59 2019<br>
> [1225]PETSC ERROR: #2 DSSolve() line 586 in /global/u1/j/jczhang/petsc/arch-cray-xc40-knl-opt/externalpackages/git.slepc/src/sys/classes/ds/interface/dsops.c<br>
> [1225]PETSC ERROR: #3 EPSSolve_KrylovSchur_Symm() line 55 in /global/u1/j/jczhang/petsc/arch-cray-xc40-knl-opt/externalpackages/git.slepc/src/eps/impls/krylov/krylovschur/ks-symm.c<br>
> [1225]PETSC ERROR: #4 EPSSolve() line 149 in /global/u1/j/jczhang/petsc/arch-cray-xc40-knl-opt/externalpackages/git.slepc/src/eps/interface/epssolve.c<br>
> [240]PETSC ERROR: #2 DSSolve() line 586 in /global/u1/j/jczhang/petsc/arch-cray-xc40-knl-opt/externalpackages/git.slepc/src/sys/classes/ds/interface/dsops.c<br>
> [240]PETSC ERROR: #3 EPSSolve_KrylovSchur_Symm() line 55 in /global/u1/j/jczhang/petsc/arch-cray-xc40-knl-opt/externalpackages/git.slepc/src/eps/impls/krylov/krylovschur/ks-symm.c<br>
> [240]PETSC ERROR: #4 EPSSolve() line 149 in /global/u1/j/jczhang/petsc/arch-cray-xc40-knl-opt/externalpackages/git.slepc/src/eps/interface/epssolve.c<br>
> <br>
> --Junchao Zhang<br>
> <br>
> <br>
> On Fri, Jun 28, 2019 at 4:02 AM Ale Foggia <<a href="mailto:amfoggia@gmail.com" target="_blank">amfoggia@gmail.com</a>> wrote:<br>
> Junchao,<br>
> I'm sorry for the late response.<br>
> <br>
> El mié., 26 jun. 2019 a las 16:39, Zhang, Junchao (<<a href="mailto:jczhang@mcs.anl.gov" target="_blank">jczhang@mcs.anl.gov</a>>) escribió:<br>
> Ale,<br>
> The job got a chance to run but failed with out-of-memory, "Some of your processes may have been killed by the cgroup out-of-memory handler."<br>
> <br>
> I mentioned that I used 1024 nodes and 32 processes on each node because the application needs a lot of memory. I think that for a system of size 38, one needs above 256 nodes for sure (assuming only 32 procs per node). I would try with 512 if it's possible.<br>
> <br>
> I also tried with 128 core with ./main.x 2 ... and got a weird error message "The size of the basis has to be at least equal to the number of MPI processes used."<br>
> <br>
> The error comes from the fact that you put a system size of only 2 which is too small.<br>
> I can also see the problem in the assembly with system sizes smaller than 38, so you can try with like 30 (for which I also have a log). In that case I run with 64 nodes and 32 processes per node. I think the problem may also fit in 32 nodes.<br>
> <br>
> --Junchao Zhang<br>
> <br>
> <br>
> On Tue, Jun 25, 2019 at 11:24 PM Junchao Zhang <<a href="mailto:jczhang@mcs.anl.gov" target="_blank">jczhang@mcs.anl.gov</a>> wrote:<br>
> Ale,<br>
> I successfully built your code and submitted a job to the NERSC Cori machine requiring 32768 KNL cores and one and a half hours. It is estimated to run in 3 days. If you also observed the same problem with less cores, what is your input arguments? Currently,
I use what in your log file, ./main.x 38 -nn -j1 1.0 -d1 1.0 -eps_type krylovschur -eps_tol 1e-9 -log_view<br>
> The smaller the better. Thanks.<br>
> --Junchao Zhang<br>
> <br>
> <br>
> On Mon, Jun 24, 2019 at 6:20 AM Ale Foggia <<a href="mailto:amfoggia@gmail.com" target="_blank">amfoggia@gmail.com</a>> wrote:<br>
> Yes, I used KNL nodes. I you can perform the test would be great. Could it be that I'm not using the correct configuration of the KNL nodes? These are the environment variables I set:<br>
> MKL_NUM_THREADS=1<br>
> OMP_NUM_THREADS=1<br>
> KMP_HW_SUBSET=1t<br>
> KMP_AFFINITY=compact<br>
> I_MPI_PIN_DOMAIN=socket<br>
> I_MPI_PIN_PROCESSOR_LIST=0-63<br>
> MKL_DYNAMIC=0<br>
> <br>
> The code is in <a href="https://github.com/amfoggia/LSQuantumED" rel="noreferrer" target="_blank">
https://github.com/amfoggia/LSQuantumED</a> and it has a readme to compile it and run it. When I run the test I used only 32 processors per node, and I used 1024 nodes in total, and it's for nspins=38.<br>
> Thank you<br>
> <br>
> El vie., 21 jun. 2019 a las 20:03, Zhang, Junchao (<<a href="mailto:jczhang@mcs.anl.gov" target="_blank">jczhang@mcs.anl.gov</a>>) escribió:<br>
> Ale,<br>
> Did you use Intel KNL nodes? Mr. Hong (cc'ed) did experiments on KNL nodes one year ago. He used 32768 processors and called MatAssemblyEnd 118 times and it used only 1.5 seconds in total. So I guess something was wrong with your test. If you can share
your code, I can have a test on our machine to see how it goes.<br>
> Thanks.<br>
> --Junchao Zhang<br>
> <br>
> <br>
> On Fri, Jun 21, 2019 at 11:00 AM Junchao Zhang <<a href="mailto:jczhang@mcs.anl.gov" target="_blank">jczhang@mcs.anl.gov</a>> wrote:<br>
> MatAssembly was called once (in stage 5) and cost 2.5% of the total time. Look at stage 5. It says MatAssemblyBegin calls BuildTwoSidedF, which does global synchronization. The high max/min ratio means load imbalance. What I do not understand is MatAssemblyEnd.
The ratio is 1.0. It means processors are already synchronized. With 32768 processors, there are 1.2e+06 messages with average length 1.9e+06 bytes. So each processor sends 36 (1.2e+06/32768) ~2MB messages and it takes 54 seconds. Another chance is the reduction
at MatAssemblyEnd. I don't know why it needs 8 reductions. In my mind, one is enough. I need to look at the code.<br>
> <br>
> Summary of Stages: ----- Time ------ ----- Flop ------ --- Messages --- -- Message Lengths -- -- Reductions --<br>
> Avg %Total Avg %Total Count %Total Avg %Total Count %Total
<br>
> 0: Main Stage: 8.5045e+02 13.0% 3.0633e+15 14.0% 8.196e+07 13.1% 7.768e+06 13.1% 2.530e+02 13.0%
<br>
> 1: Create Basis: 7.9234e-02 0.0% 0.0000e+00 0.0% 0.000e+00 0.0% 0.000e+00 0.0% 0.000e+00 0.0%
<br>
> 2: Create Lattice: 8.3944e-05 0.0% 0.0000e+00 0.0% 0.000e+00 0.0% 0.000e+00 0.0% 0.000e+00 0.0%
<br>
> 3: Create Hamilt: 1.0694e+02 1.6% 0.0000e+00 0.0% 0.000e+00 0.0% 0.000e+00 0.0% 2.000e+00 0.1%
<br>
> 5: Offdiag: 1.6525e+02 2.5% 0.0000e+00 0.0% 1.188e+06 0.2% 1.942e+06 0.0% 8.000e+00 0.4%
<br>
> 6: Phys quantities: 5.4045e+03 82.8% 1.8866e+16 86.0% 5.417e+08 86.7% 7.768e+06 86.8% 1.674e+03 86.1%
<br>
> <br>
> --- Event Stage 5: Offdiag<br>
> BuildTwoSidedF 1 1.0 7.1565e+01 148448.9 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 1 0 0 0 0 28 0 0 0 0 0<br>
> MatAssemblyBegin 1 1.0 7.1565e+01 127783.7 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 1 0 0 0 0 28 0 0 0 0 0<br>
> MatAssemblyEnd 1 1.0 5.3762e+01 1.0 0.00e+00 0.0 1.2e+06 1.9e+06 8.0e+00 1 0 0 0 0 33 0100100100 0<br>
> VecSet 1 1.0 7.5533e-02 9.0 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0<br>
> <br>
> <br>
> --Junchao Zhang<br>
> <br>
> <br>
> On Fri, Jun 21, 2019 at 10:34 AM Smith, Barry F. <<a href="mailto:bsmith@mcs.anl.gov" target="_blank">bsmith@mcs.anl.gov</a>> wrote:<br>
> <br>
> The load balance is definitely out of whack. <br>
> <br>
> <br>
> <br>
> BuildTwoSidedF 1 1.0 1.6722e-0241.0 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0<br>
> MatMult 138 1.0 2.6604e+02 7.4 3.19e+10 2.1 8.2e+07 7.8e+06 0.0e+00 2 4 13 13 0 15 25100100 0 2935476<br>
> MatAssemblyBegin 1 1.0 1.6807e-0236.1 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0<br>
> MatAssemblyEnd 1 1.0 3.5680e-01 3.9 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0<br>
> VecNorm 2 1.0 4.4252e+0174.8 1.73e+07 1.0 0.0e+00 0.0e+00 2.0e+00 1 0 0 0 0 5 0 0 0 1 12780<br>
> VecCopy 6 1.0 6.5655e-02 2.6 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0<br>
> VecAXPY 2 1.0 1.3793e-02 2.7 1.73e+07 1.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 41000838<br>
> VecScatterBegin 138 1.0 1.1653e+0285.8 0.00e+00 0.0 8.2e+07 7.8e+06 0.0e+00 1 0 13 13 0 4 0100100 0 0<br>
> VecScatterEnd 138 1.0 1.3653e+0222.4 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 1 0 0 0 0 4 0 0 0 0 0<br>
> VecSetRandom 1 1.0 9.6668e-01 2.2 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0<br>
> <br>
> Note that VecCopy/AXPY/SetRandom which are all embarrassingly parallel have a balance ratio above 2 which means some processes have more than twice the work of others. Meanwhile the ratio for anything with communication is extremely in balanced, some processes
get to the synchronization point well before other processes. <br>
> <br>
> The first thing I would do is worry about the load imbalance, what is its cause? is it one process with much less work than others (not great but not terrible) or is it one process with much more work then the others (terrible) or something in between. I
think once you get a handle on the load balance the rest may fall into place, otherwise we still have some exploring to do. This is not expected behavior for a good machine with a good network and a well balanced job. After you understand the load balancing
you may need to use one of the parallel performance visualization tools to see why the synchronization is out of whack.<br>
> <br>
> Good luck<br>
> <br>
> Barry<br>
> <br>
> <br>
> > On Jun 21, 2019, at 9:27 AM, Ale Foggia <<a href="mailto:amfoggia@gmail.com" target="_blank">amfoggia@gmail.com</a>> wrote:<br>
> > <br>
> > I'm sending one with a bit less time.<br>
> > I'm timing the functions also with std::chronos and for the case of 180 seconds the program runs out of memory (and crushes) before the PETSc log gets to be printed, so I know the time only from my function. Anyway, in every case, the times between std::chronos
and the PETSc log match.<br>
> > <br>
> > (The large times are in part "4b- Building offdiagonal part" or "Event Stage 5: Offdiag").<br>
> > <br>
> > El vie., 21 jun. 2019 a las 16:09, Zhang, Junchao (<<a href="mailto:jczhang@mcs.anl.gov" target="_blank">jczhang@mcs.anl.gov</a>>) escribió:<br>
> > <br>
> > <br>
> > On Fri, Jun 21, 2019 at 8:07 AM Ale Foggia <<a href="mailto:amfoggia@gmail.com" target="_blank">amfoggia@gmail.com</a>> wrote:<br>
> > Thanks both of you for your answers,<br>
> > <br>
> > El jue., 20 jun. 2019 a las 22:20, Smith, Barry F. (<<a href="mailto:bsmith@mcs.anl.gov" target="_blank">bsmith@mcs.anl.gov</a>>) escribió:<br>
> > <br>
> > Note that this is a one time cost if the nonzero structure of the matrix stays the same. It will not happen in future MatAssemblies.<br>
> > <br>
> > > On Jun 20, 2019, at 3:16 PM, Zhang, Junchao via petsc-users <<a href="mailto:petsc-users@mcs.anl.gov" target="_blank">petsc-users@mcs.anl.gov</a>> wrote:<br>
> > > <br>
> > > Those messages were used to build MatMult communication pattern for the matrix. They were not part of the matrix entries-passing you imagined, but indeed happened in MatAssemblyEnd. If you want to make sure processors do not set remote entries, you can
use MatSetOption(A,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE), which will generate an error when an off-proc entry is set.<br>
> > <br>
> > I started being concerned about this when I saw that the assembly was taking a few hundreds of seconds in my code, like 180 seconds, which for me it's a considerable time. Do you think (or maybe you need more information to answer this) that this time is
"reasonable" for communicating the pattern for the matrix? I already checked that I'm not setting any remote entries.
<br>
> > It is not reasonable. Could you send log view of that test with 180 seconds MatAssembly?<br>
> > <br>
> > Also I see (in my code) that even if there are no messages being passed during MatAssemblyBegin, it is taking time and the "ratio" is very big.<br>
> > <br>
> > > <br>
> > > <br>
> > > --Junchao Zhang<br>
> > > <br>
> > > <br>
> > > On Thu, Jun 20, 2019 at 4:13 AM Ale Foggia via petsc-users <<a href="mailto:petsc-users@mcs.anl.gov" target="_blank">petsc-users@mcs.anl.gov</a>> wrote:<br>
> > > Hello all!<br>
> > > <br>
> > > During the conference I showed you a problem happening during MatAssemblyEnd in a particular code that I have. Now, I tried the same with a simple code (a symmetric problem corresponding to the Laplacian operator in 1D, from the SLEPc Hands-On exercises).
As I understand (and please, correct me if I'm wrong), in this case the elements of the matrix are computed locally by each process so there should not be any communication during the assembly. However, in the log I get that there are messages being passed.
Also, the number of messages changes with the number of processes used and the size of the matrix. Could you please help me understand this?<br>
> > > <br>
> > > I attach the code I used and the log I get for a small problem.<br>
> > > <br>
> > > Cheers,<br>
> > > Ale<br>
> > > <br>
> > <br>
> > <log.txt><br>
> <br>
<br>
</blockquote>
</div>
</div>
</blockquote></div>