[petsc-users] Communication during MatAssemblyEnd

Mills, Richard Tran rtmills at anl.gov
Mon Jun 24 19:32:28 CDT 2019 

Hi Ale,

I don't know if this has anything to do with the strange performance you are seeing, but I notice that some of your Intel MPI settings are inconsistent and I'm not sure what you are intending. You have specified a value for I_MPI_PIN_DOMAIN and also a value for I_MPI_PIN_PROCESSOR_LIST. You can specify the domains that each MPI rank is pinned to by specifying one of these, but not both. According to what I found by a Google search for the Intel MPI documentation, if you specify both, it is I_MPI_PIN_DOMAIN that gets used.

A KNL node has only one socket per node (though it can be made to appear as if it has four virtual sockets if booted into SNC-4 mode, which doesn't seem to be a popular usage model). If you run with "I_MPI_PIN_DOMAIN=socket", then each of your MPI ranks on a node will be "pinned" to a domain that consists of all of the CPU cores available in a socket, so the MPI processes aren't really pinned at all, and can be migrated around by the system. You may see better performance by having your MPI implementation pin processes to more restrictive domains. I suggest first trying

  I_MPI_PIN_DOMAIN=auto:compact

which I believe is the default Intel MPI behavior. This will create domains by dividing the number of logical CPUs (# of hardware threads) by the number of MPI processes being started on the node, and they will be "compact" in the sense that domain members will be as "close" to each other as possible in terms of sharing cores, caches, etc.

I think that Intel MPI is ignoring your I_MPI_PIN_PROCESSOR_LIST setting, but I'll go ahead and point out that, if you are running with 64-core KNL nodes, there are 4*64 = 256 logical processors available, because each core supports four hardware threads. Asking for logical processors 0-63 might not actually be using all of the cores, as, depending on the BIOS numbering (which can be arbitrary), some of these logical processors may actually be hardware threads that share the same core.

Best regards,
Richard

On 6/24/19 4:19 AM, Ale Foggia via petsc-users wrote:
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:
MKL_NUM_THREADS=1
OMP_NUM_THREADS=1
KMP_HW_SUBSET=1t
KMP_AFFINITY=compact
I_MPI_PIN_DOMAIN=socket
I_MPI_PIN_PROCESSOR_LIST=0-63
MKL_DYNAMIC=0

The code is in https://github.com/amfoggia/LSQuantumED 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.
Thank you

El vie., 21 jun. 2019 a las 20:03, Zhang, Junchao (<jczhang at mcs.anl.gov<mailto:jczhang at mcs.anl.gov>>) escribió:
Ale,
  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.
 Thanks.
--Junchao Zhang


On Fri, Jun 21, 2019 at 11:00 AM Junchao Zhang <jczhang at mcs.anl.gov<mailto:jczhang at mcs.anl.gov>> wrote:
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.

Summary of Stages:   ----- Time ------  ----- Flop ------  --- Messages ---  -- Message Lengths --  -- Reductions --
                        Avg     %Total     Avg     %Total    Count   %Total     Avg         %Total    Count   %Total
 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%
 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%
 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%
 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%
 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%
 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%

--- Event Stage 5: Offdiag
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
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
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
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


--Junchao Zhang


On Fri, Jun 21, 2019 at 10:34 AM Smith, Barry F. <bsmith at mcs.anl.gov<mailto:bsmith at mcs.anl.gov>> wrote:

   The load balance is definitely out of whack.



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
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
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
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
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
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
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
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
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
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

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.

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.

   Good luck

  Barry


> On Jun 21, 2019, at 9:27 AM, Ale Foggia <amfoggia at gmail.com<mailto:amfoggia at gmail.com>> wrote:
>
> I'm sending one with a bit less time.
> 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.
>
> (The large times are in part "4b- Building offdiagonal part" or "Event Stage 5: Offdiag").
>
> El vie., 21 jun. 2019 a las 16:09, Zhang, Junchao (<jczhang at mcs.anl.gov<mailto:jczhang at mcs.anl.gov>>) escribió:
>
>
> On Fri, Jun 21, 2019 at 8:07 AM Ale Foggia <amfoggia at gmail.com<mailto:amfoggia at gmail.com>> wrote:
> Thanks both of you for your answers,
>
> El jue., 20 jun. 2019 a las 22:20, Smith, Barry F. (<bsmith at mcs.anl.gov<mailto:bsmith at mcs.anl.gov>>) escribió:
>
>   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.
>
> > On Jun 20, 2019, at 3:16 PM, Zhang, Junchao via petsc-users <petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov>> wrote:
> >
> > 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.
>
> 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.
> It is not reasonable. Could you send log view of that test with 180 seconds MatAssembly?
>
> 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.
>
> >
> >
> > --Junchao Zhang
> >
> >
> > On Thu, Jun 20, 2019 at 4:13 AM Ale Foggia via petsc-users <petsc-users at mcs.anl.gov<mailto:petsc-users at mcs.anl.gov>> wrote:
> > Hello all!
> >
> > 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?
> >
> > I attach the code I used and the log I get for a small problem.
> >
> > Cheers,
> > Ale
> >
>
> <log.txt>


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20190625/2c3503f9/attachment-0001.html>

More information about the petsc-users mailing list