[petsc-users] large PetscCommDuplicate overhead
Barry Smith
bsmith at mcs.anl.gov
Thu Oct 6 10:55:13 CDT 2016
Matt,
Thanks for this information. It sure looks like there is something seriously wrong with the MPI_Attr_get() on the cray for non-root process. Does any PETSc developer have access to such a machine? We need to write a test program that just calls MPI_Attr_get a bunch of times (no PETSc) to see if we can reproduce the problem and report it to Cray.
Barry
On Oct 6, 2016, at 10:45 AM, Matthew Overholt <overholt at capesim.com> wrote:
>
>
> Matthew and Barry,
>
> 1) I did a direct measurement of PetscCommDuplicate() time by tracing just
> that call (using CrayPat), and confirmed the sampling results. For 8
> processes (n=8), tracing counted a total of 101 calls, taking ~0 time on the
> root process but taking 11.78 seconds (6.3% of 188 total seconds) on each of
> the other 7 processes. For 16 processes (n=16, still only 1 node), tracing
> counted 102 total calls for a total of 18.42 seconds (13.2% of 139.6 total
> seconds) on every process except the root.
>
> 2) Copied below is a section of the log view for the first two solutions for
> n=2, which shows the same calls as for n=8. (I can send the entire log files
> if desired.) In each case I count about 44 PCD calls per process during
> initialization and meshing, 7 calls during setup, 9 calls for the first
> solution, then 3 calls for each subsequent solution (fixed-point iteration),
> and 3 calls to write out the solution, for 75 total.
>
> 3) I would expect that the administrators of this machine have configured
> PETSc appropriately. I am using their current default install, which is
> 3.7.2.
> https://www.nersc.gov/users/software/programming-libraries/math-libraries/pe
> tsc/
>
> 4) Yes, I just gave the MUMPS time as a comparison.
>
> 5) As to where it is spending time, perhaps the timing results in the log
> files will be helpful. The "Solution took ..." printouts give the total
> solution time for that iteration, the others are incremental times. (As an
> aside, I have been wondering why the solution times do not scale well with
> process count, even though that work is entirely done in parallel PETSc
> routines.)
>
> Thanks,
> Matt Overholt
>
>
> ********** -log_view -info results for n=2 : the first solution and
> subsequent fixed-point iteration ***********
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850688
> -2080374779
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850688
> -2080374779
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850688
> -2080374781
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> Matrix setup took 0.108 s
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850688
> -2080374779
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850688
> -2080374781
> KSP PC setup took 0.079 s
> [0] VecAssemblyBegin_MPI(): Stash has 0 entries, uses 0 mallocs.
> [0] VecAssemblyBegin_MPI(): Block-Stash has 0 entries, uses 0 mallocs.
> [0] MatStashScatterBegin_Ref(): No of messages: 0
> [0] MatAssemblyBegin_MPIAIJ(): Stash has 0 entries, uses 0 mallocs.
> [1] MatStashScatterBegin_Ref(): No of messages: 1
> [1] MatStashScatterBegin_Ref(): Mesg_to: 0: size: 7713792 bytes
> [1] MatAssemblyBegin_MPIAIJ(): Stash has 482112 entries, uses 5 mallocs.
> [1] MatAssemblyEnd_SeqAIJ(): Matrix size: 599214 X 599214; storage space:
> 1050106 unneeded,15128672 used
> [1] MatAssemblyEnd_SeqAIJ(): Number of mallocs during MatSetValues() is 0
> [1] MatAssemblyEnd_SeqAIJ(): Maximum nonzeros in any row is 27
> [1] MatCheckCompressedRow(): Found the ratio (num_zerorows 0)/(num_localrows
> 599214) < 0.6. Do not use CompressedRow routines.
> [1] MatSeqAIJCheckInode(): Found 599214 nodes out of 599214 rows. Not using
> Inode routines
> [0] MatAssemblyEnd_SeqAIJ(): Matrix size: 621594 X 621594; storage space:
> 1237634 unneeded,15545404 used
> [0] MatAssemblyEnd_SeqAIJ(): Number of mallocs during MatSetValues() is 0
> [0] MatAssemblyEnd_SeqAIJ(): Maximum nonzeros in any row is 27
> [0] MatCheckCompressedRow(): Found the ratio (num_zerorows 0)/(num_localrows
> 621594) < 0.6. Do not use CompressedRow routines.
> [0] MatSeqAIJCheckInode(): Found 621594 nodes out of 621594 rows. Not using
> Inode routines
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [0] VecScatterCreateCommon_PtoS(): Using MPI_Alltoallv() for scatter
> [0] VecScatterCreateCommon_PtoS(): Using blocksize 1 scatter
> [0] VecScatterCreate(): General case: MPI to Seq
> [1] MatAssemblyEnd_SeqAIJ(): Matrix size: 599214 X 15700; storage space:
> 5257543 unneeded,136718 used
> [1] MatAssemblyEnd_SeqAIJ(): Number of mallocs during MatSetValues() is 89
> [1] MatAssemblyEnd_SeqAIJ(): Maximum nonzeros in any row is 19
> [1] MatCheckCompressedRow(): Found the ratio (num_zerorows
> 582718)/(num_localrows 599214) > 0.6. Use CompressedRow routines.
> [0] MatAssemblyEnd_SeqAIJ(): Matrix size: 621594 X 16496; storage space:
> 5464978 unneeded,136718 used
> [0] MatAssemblyEnd_SeqAIJ(): Number of mallocs during MatSetValues() is 490
> [0] MatAssemblyEnd_SeqAIJ(): Maximum nonzeros in any row is 16
> [0] MatCheckCompressedRow(): Found the ratio (num_zerorows
> 605894)/(num_localrows 621594) > 0.6. Use CompressedRow routines.
> K and q SetValues took 26.426 s
> [0] PCSetUp(): Setting up PC for first time
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [0] VecScatterCreate(): Special case: processor zero gets entire parallel
> vector, rest get none
> ** Max-trans not allowed because matrix is distributed
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [0] PCSetUp(): Leaving PC with identical preconditioner since operator is
> unchanged
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [0] VecScatterCreateCommon_PtoS(): Using MPI_Alltoallv() for scatter
> [0] VecScatterCreateCommon_PtoS(): Using blocksize 1 scatter
> [0] VecScatterCreate(): General case: Seq to MPI
> [1] VecScatterCreate(): General case: Seq to MPI
> Solution took 102.21 s
>
> NL iteration 0: delta = 32.0488 67.6279.
> Error delta calc took 0.045 s
> Node and Element temps update took 0.017 s
> [0] VecAssemblyBegin_MPI(): Stash has 0 entries, uses 0 mallocs.
> [0] VecAssemblyBegin_MPI(): Block-Stash has 0 entries, uses 0 mallocs.
> [0] MatStashScatterBegin_Ref(): No of messages: 0
> [0] MatAssemblyBegin_MPIAIJ(): Stash has 0 entries, uses 0 mallocs.
> [1] MatStashScatterBegin_Ref(): No of messages: 1
> [1] MatStashScatterBegin_Ref(): Mesg_to: 0: size: 7713792 bytes
> [1] MatAssemblyBegin_MPIAIJ(): Stash has 482112 entries, uses 0 mallocs.
> [1] MatAssemblyEnd_SeqAIJ(): Matrix size: 599214 X 599214; storage space: 0
> unneeded,15128672 used
> [1] MatAssemblyEnd_SeqAIJ(): Number of mallocs during MatSetValues() is 0
> [1] MatAssemblyEnd_SeqAIJ(): Maximum nonzeros in any row is 27
> [1] MatCheckCompressedRow(): Found the ratio (num_zerorows 0)/(num_localrows
> 599214) < 0.6. Do not use CompressedRow routines.
> [0] MatAssemblyEnd_SeqAIJ(): Matrix size: 621594 X 621594; storage space: 0
> unneeded,15545404 used
> [0] MatAssemblyEnd_SeqAIJ(): Number of mallocs during MatSetValues() is 0
> [0] MatAssemblyEnd_SeqAIJ(): Maximum nonzeros in any row is 27
> [0] MatCheckCompressedRow(): Found the ratio (num_zerorows 0)/(num_localrows
> 621594) < 0.6. Do not use CompressedRow routines.
> [1] MatAssemblyEnd_SeqAIJ(): Matrix size: 599214 X 15700; storage space: 0
> unneeded,136718 used
> [1] MatAssemblyEnd_SeqAIJ(): Number of mallocs during MatSetValues() is 0
> [1] MatAssemblyEnd_SeqAIJ(): Maximum nonzeros in any row is 19
> [1] MatCheckCompressedRow(): Found the ratio (num_zerorows
> 582718)/(num_localrows 599214) > 0.6. Use CompressedRow routines.
> [0] MatAssemblyEnd_SeqAIJ(): Matrix size: 621594 X 16496; storage space: 0
> unneeded,136718 used
> [0] MatAssemblyEnd_SeqAIJ(): Number of mallocs during MatSetValues() is 0
> [0] MatAssemblyEnd_SeqAIJ(): Maximum nonzeros in any row is 16
> [0] MatCheckCompressedRow(): Found the ratio (num_zerorows
> 605894)/(num_localrows 621594) > 0.6. Use CompressedRow routines.
> K and q SetValues took 2.366 s
> [0] PCSetUp(): Setting up PC with same nonzero pattern
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [0] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374780
> [1] PetscCommDuplicate(): Using internal PETSc communicator 1140850689
> -2080374782
> [0] VecScatterCreateCommon_PtoS(): Using MPI_Alltoallv() for scatter
> [0] VecScatterCreateCommon_PtoS(): Using blocksize 1 scatter
> [0] VecScatterCreate(): General case: Seq to MPI
> [1] VecScatterCreate(): General case: Seq to MPI
> Solution took 82.156 s
>
> -----Original Message-----
> From: Barry Smith [mailto:bsmith at mcs.anl.gov]
> Sent: Wednesday, October 05, 2016 4:42 PM
> To: overholt at capesim.com
> Cc: petsc-users at mcs.anl.gov
> Subject: Re: [petsc-users] large PetscCommDuplicate overhead
>
>
>> On Oct 5, 2016, at 2:30 PM, Matthew Overholt <overholt at capesim.com> wrote:
>>
>> Hi Petsc-Users,
>>
>> I am trying to understand an issue where PetscCommDuplicate() calls are
> taking an increasing percentage of time as I run a fixed-sized problem on
> more processes.
>>
>> I am using the FEM to solve the steady-state heat transfer equation (K.x =
> q) using a PC direct solver, like MUMPS.
>>
>> I am running on the NERSC Cray X30, which has two Xeon's per node with 12
> cores each, and profiling the code using CrayPat sampling.
>>
>> On a typical problem (1E+6 finite elements), running on a single node:
>> -for 2 cores (1 on each Xeon), about 1% of time is PetscCommDuplicate (on
> process 1, but on the root it is less), and (for reference) 9% of total time
> is for MUMPS.
>> -for 8 cores (4 on each Xeon), over 6% of time is PetscCommDuplicate (on
> every process except the root, where it is <1%), and 9-10% of total time is
> for MUMPS.
>
> What does PetscCommDuplicate() have to do with MUMPS? Nothing at all, you
> are just giving its time for comparison?
>
>>
>> What is the large PetscCommDuplicate time connected to, an increasing
> number of messages (tags)? Would using fewer MatSetValues() and
> VecSetValues() calls (with longer message lengths) alleviate this?
>
> No PetscCommDuplicate won't increate with more messages or calls to
> XXXSetValues(). PetscCommDuplicate() is only called essentially on the
> creation of new PETSc objects. It should also be fast since it basically
> needs to do just a MPI_Attr_get(). With more processes but the same problem
> size and code there should be pretty much the same number of objects
> created.
>
> PetscSpinlockLock() does nothing if you are not using threads so it won't
> take any time.
>
> Is there a way to see where it is spending its time inside the
> PetscCommDuplicate()? Perhaps the Cray MPI_Attr_get() has issues.
>
> Barry
>
>
>
>
>
>
>>
>> For reference, the PETSc calling sequence in the code is as follows.
>> // Create the solution and RHS vectors
>> ierr = VecCreate(petscData->mpicomm,&mesh->hpx);
>> ierr = PetscObjectSetName((PetscObject) mesh->hpx, "Solution");
>> ierr = VecSetSizes(mesh->hpx,mesh->lxN,mesh->neqns); // size = # of
> equations; distribution to match mesh
>> ierr = VecSetFromOptions(mesh->hpx); // allow run time options
>> ierr = VecDuplicate(mesh->hpx,&q); // create the RHS vector
>> // Create the stiffnexx matrix
>> ierr = MatCreate(petscData->mpicomm,&K);
>> ierr = MatSetSizes(K,mesh->lxN,mesh->lxN,mesh->neqns,mesh->neqns);
>> ierr = MatSetType(K,MATAIJ); // default sparse type
>> // Do preallocation
>> ierr = MatMPIAIJSetPreallocation(K,d_nz,NULL,o_nz,NULL);
>> ierr = MatSeqAIJSetPreallocation(K,d_nz,NULL);
>> ierr = MatSetOption(K,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);
>> ierr = MatSetUp(K);
>> // Create and set up the KSP context as a PreConditioner Only (Direct)
> Solution
>> ierr = KSPCreate(petscData->mpicomm,&ksp);
>> ierr = KSPSetOperators(ksp,K,K);
>> ierr = KSPSetType(ksp,KSPPREONLY);
>> // Set the temperature vector
>> ierr = VecSet(mesh->hpx,mesh->Tmin);
>> // Set the default PC method as MUMPS
>> ierr = KSPGetPC(ksp,&pc); // extract the preconditioner
>> ierr = PCSetType(pc,PCLU); // set pc options
>> ierr = PCFactorSetMatSolverPackage(pc,MATSOLVERMUMPS);
>> ierr = KSPSetFromOptions(ksp);
>>
>> // Set the values for the K matrix and q vector
>> // which involves a lot of these calls
>> ierr = MatSetValues(K,mrows,idxm,ncols,idxn,pKe,ADD_VALUES); //
> 1 call per matrix row (equation)
>> ierr = VecSetValues(q,nqe,ixn,pqe,ADD_VALUES); // 1 call per
> element
>> ierr = VecAssemblyBegin(q);
>> ierr = MatAssemblyBegin(K,MAT_FINAL_ASSEMBLY);
>> ierr = VecAssemblyEnd(q);
>> ierr = MatAssemblyEnd(K,MAT_FINAL_ASSEMBLY);
>>
>> // Solve //////////////////////////////////////
>> ierr = KSPSolve(ksp,q,mesh->hpx);
>> ...
>> *Note that the code evenly divides the finite elements over the total
> number of processors, and I am using ghosting of the FE vertices vector to
> handle the vertices that are needed on more than 1 process.
>>
>> Thanks in advance for your help,
>> Matt Overholt
>> CapeSym, Inc.
>>
>> Virus-free. www.avast.com
>
>
> ---
> This email has been checked for viruses by Avast antivirus software.
> https://www.avast.com/antivirus
>
More information about the petsc-users
mailing list