profiling PETSc code
Barry Smith
bsmith at mcs.anl.gov
Tue Aug 15 14:56:34 CDT 2006
Please send the entire -info output as an attachment to me. (Not
in the email) I'll study it in more detail.
Barry
On Tue, 15 Aug 2006, Matt Funk wrote:
> On Tuesday 15 August 2006 11:52, Barry Smith wrote:
>>> MatSolve 16000 1.0 1.1285e+02 1.4 1.50e+08 1.4 0.0e+00 0.0e+00
>>
>> ^^^^^
>> balance
>>
>> Hmmm, I would guess that the matrix entries are not so well balanced?
>> One process takes 1.4 times as long for the triangular solves as the other
>> so either one matrix has many more entries or one processor is slower then
>> the other.
>>
>> Barry
>
> Well it would seem that way at first, but i don't know how that could be since
> i allocate an exactly equal amount of points on both processor (see previous
> email).
> Further i used the -mat_view_info option. Here is what it gives me:
>
> ...
> Matrix Object:
> type=mpiaij, rows=119808, cols=119808
> [1] PetscCommDuplicateUsing internal PETSc communicator 91 168
> [1] PetscCommDuplicateUsing internal PETSc communicator 91 168
> ...
>
> ...
> Matrix Object:
> type=seqaij, rows=59904, cols=59904
> total: nonzeros=407400, allocated nonzeros=407400
> not using I-node routines
> Matrix Object:
> type=seqaij, rows=59904, cols=59904
> total: nonzeros=407400, allocated nonzeros=407400
> not using I-node routines
> ...
>
> So to me it look s well split up. Is there anything else that somebody can
> think of. The machine i am running on is all same processors.
>
> By the way, i am not sure if the '[1] PetscCommDuplicateUsing internal PETSc
> communicator 91 168' is something i need to worry about??
>
> mat
>
>
>
>
>
>
>
>
>>
>> On Tue, 15 Aug 2006, Matt Funk wrote:
>>> Hi Matt,
>>>
>>> sorry for the delay since the last email, but there were some other
>>> things i needed to do.
>>>
>>> Anyway, I hope that maybe I can get some more help from you guys with
>>> respect to the loadimbalance problem i have. Here is the situtation:
>>> I run my code on 2 procs. I profile my KSPSolve call and here is what i
>>> get:
>>>
>>> ...
>>>
>>> --- Event Stage 2: Stage 2 of ChomboPetscInterface
>>>
>>> VecDot 20000 1.0 1.9575e+01 2.0 2.39e+08 2.0 0.0e+00 0.0e+00
>>> 2.0e+04 2 8 0 0 56 7 8 0 0 56 245
>>> VecNorm 16000 1.0 4.0559e+01 3.2 1.53e+08 3.2 0.0e+00 0.0e+00
>>> 1.6e+04 3 7 0 0 44 13 7 0 0 44 95
>>> VecCopy 4000 1.0 1.5148e+00 1.4 0.00e+00 0.0 0.0e+00 0.0e+00
>>> 0.0e+00 0 0 0 0 0 1 0 0 0 0 0
>>> VecSet 16000 1.0 3.1937e+00 1.8 0.00e+00 0.0 0.0e+00 0.0e+00
>>> 0.0e+00 0 0 0 0 0 1 0 0 0 0 0
>>> VecAXPY 16000 1.0 8.2395e+00 1.4 3.22e+08 1.4 0.0e+00 0.0e+00
>>> 0.0e+00 1 7 0 0 0 3 7 0 0 0 465
>>> VecAYPX 8000 1.0 3.6370e+00 1.3 3.46e+08 1.3 0.0e+00 0.0e+00
>>> 0.0e+00 0 3 0 0 0 2 3 0 0 0 527
>>> VecScatterBegin 12000 1.0 5.7721e-01 1.1 0.00e+00 0.0 2.4e+04 2.1e+04
>>> 0.0e+00 0 0100100 0 0 0100100 0 0
>>> VecScatterEnd 12000 1.0 1.4059e+01 9.2 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
>>> MatMult 12000 1.0 5.0567e+01 1.0 1.82e+08 1.0 2.4e+04 2.1e+04
>>> 0.0e+00 5 32100100 0 25 32100100 0 361
>>> MatSolve 16000 1.0 1.1285e+02 1.4 1.50e+08 1.4 0.0e+00 0.0e+00
>>> 0.0e+00 10 43 0 0 0 47 43 0 0 0 214
>>> MatLUFactorNum 1 1.0 1.9693e-02 1.0 5.79e+07 1.1 0.0e+00 0.0e+00
>>> 0.0e+00 0 0 0 0 0 0 0 0 0 0 110
>>> MatILUFactorSym 1 1.0 7.8840e-03 1.1 0.00e+00 0.0 0.0e+00 0.0e+00
>>> 1.0e+00 0 0 0 0 0 0 0 0 0 0 0
>>> MatGetOrdering 1 1.0 1.2250e-03 1.1 0.00e+00 0.0 0.0e+00 0.0e+00
>>> 2.0e+00 0 0 0 0 0 0 0 0 0 0 0
>>> KSPSetup 1 1.0 1.1921e-06 1.2 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
>>> KSPSolve 4000 1.0 2.0419e+02 1.0 1.39e+08 1.0 2.4e+04 2.1e+04
>>> 3.6e+04 21100100100100 100100100100100 278
>>> PCSetUp 1 1.0 2.8828e-02 1.1 3.99e+07 1.1 0.0e+00 0.0e+00
>>> 3.0e+00 0 0 0 0 0 0 0 0 0 0 75
>>> PCSetUpOnBlocks 4000 1.0 3.5605e-02 1.1 3.35e+07 1.1 0.0e+00 0.0e+00
>>> 3.0e+00 0 0 0 0 0 0 0 0 0 0 61
>>> PCApply 16000 1.0 1.1661e+02 1.4 1.46e+08 1.4 0.0e+00 0.0e+00
>>> 0.0e+00 10 43 0 0 0 49 43 0 0 0 207
>>> -------------------------------------------------------------------------
>>> -----------------------------------------------
>>>
>>> ...
>>>
>>>
>>> Some things to note are the following:
>>> I allocate my vector as:
>>> VecCreateMPI(PETSC_COMM_WORLD, //communicator
>>> a_totallocal_numPoints[a_thisproc], //local points on this proc
>>> a_totalglobal_numPoints, //total number of global points
>>> &m_globalRHSVector); //the vector to be created
>>>
>>> where the vector a_totallocal_numPoints is :
>>> a_totallocal_numPoints: 59904 59904
>>>
>>> The matrix is allocated as:
>>> m_ierr = MatCreateMPIAIJ(PETSC_COMM_WORLD, //communicator
>>> a_totallocal_numPoints[a_thisproc], //total number of local rows
>>> (that is rows residing on this proc
>>> a_totallocal_numPoints[a_thisproc], //total number of columns
>>> corresponding to local part of parallel vector
>>> a_totalglobal_numPoints, //number of global rows
>>> a_totalglobal_numPoints, //number of global columns
>>> PETSC_NULL,
>>> a_NumberOfNZPointsInDiagonalMatrix,
>>> PETSC_NULL,
>>> a_NumberOfNZPointsInOffDiagonalMatrix,
>>> &m_globalMatrix);
>>>
>>> With the info option i checked and there is no extra mallocs at all.
>>> My problems setup is symmetric so it seems that everything is set up so
>>> that it should be essentially perfectly balanced. However, the numbers
>>> given above certainly do not reflect that.
>>>
>>> However, the in all other parts of my code (except the PETSc call), i get
>>> the expected, almost perfect loadbalance.
>>>
>>> Is there anything that i am overlooking? Any help is greatly appreciated.
>>>
>>> thanks
>>> mat
>>>
>>> On Wednesday 02 August 2006 16:21, Matt Funk wrote:
>>>> Hi Matt,
>>>>
>>>> It could be a bad load imbalance because i don't let PETSc decide. I
>>>> need to fix that anyway, so i think i'll try that first and then let you
>>>> know. Thanks though for the quick response and helping me to interpret
>>>> those numbers ...
>>>>
>>>>
>>>> mat
>>>>
>>>> On Wednesday 02 August 2006 15:50, Matthew Knepley wrote:
>>>>> On 8/2/06, Matt Funk <mafunk at nmsu.edu> wrote:
>>>>>> Hi Matt,
>>>>>>
>>>>>> thanks for all the help so far. The -info option is really very
>>>>>> helpful. So i think i straightened the actual errors out. However, now
>>>>>> i am back to the original question i had. That is why it takes so much
>>>>>> longer on 4 procs than on 1 proc.
>>>>>
>>>>> So you have a 1.5 load imbalance for MatMult(), which probably cascades
>>>>> to give the 133! load imbalance for VecDot(). You probably have either:
>>>>>
>>>>> 1) VERY bad laod imbalance
>>>>>
>>>>> 2) a screwed up network
>>>>>
>>>>> 3) bad contention on the network (loaded cluster)
>>>>>
>>>>> Can you help us narrow this down?
>>>>>
>>>>>
>>>>> Matt
>>>>>
>>>>>> I profiled the KSPSolve(...) as stage 2:
>>>>>>
>>>>>> For 1 proc i have:
>>>>>> --- Event Stage 2: Stage 2 of ChomboPetscInterface
>>>>>>
>>>>>> VecDot 4000 1.0 4.9158e-02 1.0 4.74e+08 1.0 0.0e+00
>>>>>> 0.0e+00 0.0e+00 0 18 0 0 0 2 18 0 0 0 474
>>>>>> VecNorm 8000 1.0 2.1798e-01 1.0 2.14e+08 1.0 0.0e+00
>>>>>> 0.0e+00 4.0e+03 1 36 0 0 28 7 36 0 0 33 214
>>>>>> VecAYPX 4000 1.0 1.3449e-01 1.0 1.73e+08 1.0 0.0e+00
>>>>>> 0.0e+00 0.0e+00 0 18 0 0 0 5 18 0 0 0 173
>>>>>> MatMult 4000 1.0 3.6004e-01 1.0 3.24e+07 1.0 0.0e+00
>>>>>> 0.0e+00 0.0e+00 1 9 0 0 0 12 9 0 0 0 32
>>>>>> MatSolve 8000 1.0 1.0620e+00 1.0 2.19e+07 1.0 0.0e+00
>>>>>> 0.0e+00 0.0e+00 3 18 0 0 0 36 18 0 0 0 22
>>>>>> KSPSolve 4000 1.0 2.8338e+00 1.0 4.52e+07 1.0 0.0e+00
>>>>>> 0.0e+00 1.2e+04 7100 0 0 84 97100 0 0100 45
>>>>>> PCApply 8000 1.0 1.1133e+00 1.0 2.09e+07 1.0 0.0e+00
>>>>>> 0.0e+00 0.0e+00 3 18 0 0 0 38 18 0 0 0 21
>>>>>>
>>>>>>
>>>>>> for 4 procs i have :
>>>>>> --- Event Stage 2: Stage 2 of ChomboPetscInterface
>>>>>>
>>>>>> VecDot 4000 1.0 3.5884e+01133.7 2.17e+07133.7 0.0e+00
>>>>>> 0.0e+00 4.0e+03 8 18 0 0 5 9 18 0 0 14 1
>>>>>> VecNorm 8000 1.0 3.4986e-01 1.3 4.43e+07 1.3 0.0e+00
>>>>>> 0.0e+00 8.0e+03 0 36 0 0 10 0 36 0 0 29 133
>>>>>> VecSet 8000 1.0 3.5024e-02 1.4 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
>>>>>> VecAYPX 4000 1.0 5.6790e-02 1.3 1.28e+08 1.3 0.0e+00
>>>>>> 0.0e+00 0.0e+00 0 18 0 0 0 0 18 0 0 0 410
>>>>>> VecScatterBegin 4000 1.0 6.0042e+01 1.4 0.00e+00 0.0 0.0e+00
>>>>>> 0.0e+00 0.0e+00 38 0 0 0 0 45 0 0 0 0 0
>>>>>> VecScatterEnd 4000 1.0 5.9364e+01 1.4 0.00e+00 0.0 0.0e+00
>>>>>> 0.0e+00 0.0e+00 37 0 0 0 0 44 0 0 0 0 0
>>>>>> MatMult 4000 1.0 1.1959e+02 1.4 3.46e+04 1.4 0.0e+00
>>>>>> 0.0e+00 0.0e+00 75 9 0 0 0 89 9 0 0 0 0
>>>>>> MatSolve 8000 1.0 2.8150e-01 1.0 2.16e+07 1.0 0.0e+00
>>>>>> 0.0e+00 0.0e+00 0 18 0 0 0 0 18 0 0 0 83
>>>>>> MatLUFactorNum 1 1.0 1.3685e-04 1.1 5.64e+06 1.1 0.0e+00
>>>>>> 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 21
>>>>>> MatILUFactorSym 1 1.0 2.3389e-04 1.2 0.00e+00 0.0 0.0e+00
>>>>>> 0.0e+00 2.0e+00 0 0 0 0 0 0 0 0 0 0 0
>>>>>> MatGetOrdering 1 1.0 9.6083e-05 1.2 0.00e+00 0.0 0.0e+00
>>>>>> 0.0e+00 2.0e+00 0 0 0 0 0 0 0 0 0 0 0
>>>>>> KSPSetup 1 1.0 2.1458e-06 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
>>>>>> KSPSolve 4000 1.0 1.2200e+02 1.0 2.63e+05 1.0 0.0e+00
>>>>>> 0.0e+00 2.8e+04 84100 0 0 34 100100 0 0100 1
>>>>>> PCSetUp 1 1.0 5.0187e-04 1.2 1.68e+06 1.2 0.0e+00
>>>>>> 0.0e+00 4.0e+00 0 0 0 0 0 0 0 0 0 0 6
>>>>>> PCSetUpOnBlocks 4000 1.0 1.2104e-02 2.2 1.34e+05 2.2 0.0e+00
>>>>>> 0.0e+00 4.0e+00 0 0 0 0 0 0 0 0 0 0 0
>>>>>> PCApply 8000 1.0 8.4254e-01 1.2 8.27e+06 1.2 0.0e+00
>>>>>> 0.0e+00 8.0e+03 1 18 0 0 10 1 18 0 0 29 28
>>>>>> ----------------------------------------------------------------------
>>>>>> - -- -----------------------------------------------
>>>>>>
>>>>>> Now if i understand it right, all these calls summarize all calls
>>>>>> between the pop and push commands. That would mean that the majority
>>>>>> of the time is spend in the MatMult and in within that the
>>>>>> VecScatterBegin and VecScatterEnd commands (if i understand it right).
>>>>>>
>>>>>> My problem size is really small. So i was wondering if the problem
>>>>>> lies in that (namely that the major time is simply spend communicating
>>>>>> between processors, or whether there is still something wrong with how
>>>>>> i wrote the code?)
>>>>>>
>>>>>>
>>>>>> thanks
>>>>>> mat
>>>>>>
>>>>>> On Tuesday 01 August 2006 18:28, Matthew Knepley wrote:
>>>>>>> On 8/1/06, Matt Funk <mafunk at nmsu.edu> wrote:
>>>>>>>> Actually the errors occur on my calls to a PETSc functions after
>>>>>>>> calling PETSCInitialize.
>>>>>>>
>>>>>>> Yes, it is the error I pointed out in the last message.
>>>>>>>
>>>>>>> Matt
>>>>>>>
>>>>>>>> mat
>
>
More information about the petsc-users
mailing list