[petsc-users] DMPlex memory problem in scaling test

Dave May dave.mayhem23 at gmail.com
Thu Oct 10 08:10:42 CDT 2019 

On Thu 10. Oct 2019 at 15:04, Matthew Knepley <knepley at gmail.com> wrote:

> On Thu, Oct 10, 2019 at 8:41 AM Dave May <dave.mayhem23 at gmail.com> wrote:
>
>> On Thu 10. Oct 2019 at 14:34, Matthew Knepley <knepley at gmail.com> wrote:
>>
>>> On Thu, Oct 10, 2019 at 8:31 AM Dave May <dave.mayhem23 at gmail.com>
>>> wrote:
>>>
>>>> On Thu, 10 Oct 2019 at 13:21, Matthew Knepley via petsc-users <
>>>> petsc-users at mcs.anl.gov> wrote:
>>>>
>>>>> On Wed, Oct 9, 2019 at 5:10 PM Danyang Su via petsc-users <
>>>>> petsc-users at mcs.anl.gov> wrote:
>>>>>
>>>>>> Dear All,
>>>>>>
>>>>>> I have a question regarding the maximum memory usage for the scaling
>>>>>> test. My code is written in Fortran with support for both structured grid
>>>>>> (DM) and unstructured grid (DMPlex). It looks like memory consumption is
>>>>>> much larger when DMPlex is used and finally causew out_of_memory problem.
>>>>>>
>>>>>> Below are some test using both structured grid and unstructured grid.
>>>>>> The memory consumption by the code is estimated based on all allocated
>>>>>> arrays and PETSc memory consumption is estimated based on
>>>>>> PetscMemoryGetMaximumUsage.
>>>>>>
>>>>>> I just wonder why the PETSc memory consumption does not decrease when
>>>>>> number of processors increases. For structured grid (scenario 7-9), the
>>>>>> memory consumption decreases as number of processors increases. However,
>>>>>> for unstructured grid case (scenario 14-16), the memory for PETSc part
>>>>>> remains unchanged. When I run a larger case, the code crashes because
>>>>>> memory is ran out. The same case works on another cluster with 480GB memory
>>>>>> per node. Does this make sense?
>>>>>>
>>>>> We would need a finer breakdown of where memory is being used. I did
>>>>> this for a paper:
>>>>>
>>>>>   https://agupubs.onlinelibrary.wiley.com/doi/full/10.1002/jgrb.50217
>>>>>
>>>>> If the subdomains, the halo sizes can overwhelm the basic storage. It
>>>>> looks like the subdomains are big here,
>>>>> but things are not totally clear to me. It would be helpful to send
>>>>> the output of -log_view for each case since
>>>>> PETSc tries to keep track of allocated memory.
>>>>>
>>>>
>>>> Matt - I'd guess that there is a sequential (non-partitioned) mesh
>>>> hanging around in memory.
>>>> Is it possible that he's created the PLEX object which is loaded
>>>> sequentially (stored and retained in memory and never released), and then
>>>> afterwards distributed?
>>>> This can never happen with the DMDA and the table verifies this.
>>>> If his code using the DMDA and DMPLEX are as identical as possible
>>>> (albeit the DM used), then a sequential mesh held in memory seems the
>>>> likely cause.
>>>>
>>>
>>> Dang it, Dave is always right.
>>>
>>> How to prevent this?
>>>
>>
>> I thought you/Lawrence/Vaclav/others... had developed and provided
>> support  for a parallel DMPLEX load via a suitably defined plex specific H5
>> mesh file.
>>
>
> We have, but these tests looked like generated meshes.
>

Great.

So would a solution to the problem be to have the user modify their code in
the follow way:
* they move the mesh gen stage into a seperate exec which they call offline
(on a fat node with lots of memory), and dump the appropriate file
* they change their existing application to simply load that file in
parallel.

If there were examples illustrating how to create the file which can be
loaded in parallel I think it would be very helpful for the user (and many
others)

Cheers
Dave


>   Thanks,
>
>     Matt
>
>
>> Since it looks like you are okay with fairly regular meshes, I would
>>> construct the
>>> coarsest mesh you can, and then use
>>>
>>>   -dm_refine <k>
>>>
>>> which is activated by DMSetFromOptions(). Make sure to call it after
>>> DMPlexDistribute(). It will regularly
>>> refine in parallel and should show good memory scaling as Dave says.
>>>
>>>   Thanks,
>>>
>>>      Matt
>>>
>>>
>>>>
>>>>>   Thanks,
>>>>>
>>>>>      Matt
>>>>>
>>>>>> scenario no. points cell type DMPLex nprocs no. nodes mem per node GB
>>>>>> solver Rank 0 memory MB Rank 0 petsc memory MB Runtime (sec)
>>>>>> 1 2121 rectangle no 40 1 200 GMRES,Hypre preconditioner 0.21 41.6
>>>>>> 2 8241 rectangle no 40 1 200 GMRES,Hypre preconditioner 0.59 51.84
>>>>>> 3 32481 rectangle no 40 1 200 GMRES,Hypre preconditioner 1.95 59.1
>>>>>> 4 128961 rectangle no 40 1 200 GMRES,Hypre preconditioner 7.05 89.71
>>>>>> 5 513921 rectangle no 40 1 200 GMRES,Hypre preconditioner 26.76
>>>>>> 110.58
>>>>>> 6 2051841 rectangle no 40 1 200 GMRES,Hypre preconditioner 104.21
>>>>>> 232.05
>>>>>> *7* *8199681* *rectangle* *no* *40* *1* *200* *GMRES,Hypre
>>>>>> preconditioner* *411.26* *703.27* *140.29*
>>>>>> *8* *8199681* *rectangle* *no* *80* *2* *200* *GMRES,Hypre
>>>>>> preconditioner* *206.6* *387.25* *62.04*
>>>>>> *9* *8199681* *rectangle* *no* *160* *4* *200* *GMRES,Hypre
>>>>>> preconditioner* *104.28* *245.3* *32.76*
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> 10 2121 triangle yes 40 1 200 GMRES,Hypre preconditioner 0.49 61.78
>>>>>> 11 15090 triangle yes 40 1 200 GMRES,Hypre preconditioner 2.32 96.61
>>>>>> 12 59847 triangle yes 40 1 200 GMRES,Hypre preconditioner 8.28 176.14
>>>>>> 13 238568 triangle yes 40 1 200 GMRES,Hypre preconditioner 31.89
>>>>>> 573.73
>>>>>> *14* *953433* *triangle* *yes* *40* *1* *200* *GMRES,Hypre
>>>>>> preconditioner* *119.23* *2102.54* *44.11*
>>>>>> *15* *953433* *triangle* *yes* *80* *2* *200* *GMRES,Hypre
>>>>>> preconditioner* *72.99* *2123.8* *24.36*
>>>>>> *16* *953433* *triangle* *yes* *160* *4* *200* *GMRES,Hypre
>>>>>> preconditioner* *48.65* *2076.25* *14.87*
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> 17 55770 prism yes 40 1 200 GMRES,Hypre preconditioner 18.46 219.39
>>>>>> 18 749814 prism yes 40 1 200 GMRES,Hypre preconditioner 149.86
>>>>>> 2412.39
>>>>>> 19 7000050 prism yes 40 to 640 1 to 16 200 GMRES,Hypre preconditioner
>>>>>> out_of_memory
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> *20* *7000050* *prism* *yes* *64* *2* *480* *GMRES,Hypre
>>>>>> preconditioner* *890.92* *17214.41*
>>>>>>
>>>>>> The error information of scenario 19 is shown below:
>>>>>>
>>>>>> kernel messages produced during job executions:
>>>>>> [Oct 9 10:41] mpiexec.hydra invoked oom-killer: gfp_mask=0x200da,
>>>>>> order=0, oom_score_adj=0
>>>>>> [  +0.010274] mpiexec.hydra cpuset=/ mems_allowed=0-1
>>>>>> [  +0.006680] CPU: 2 PID: 144904 Comm: mpiexec.hydra Tainted:
>>>>>> G           OE  ------------   3.10.0-862.14.4.el7.x86_64 #1
>>>>>> [  +0.013365] Hardware name: Lenovo ThinkSystem SD530
>>>>>> -[7X21CTO1WW]-/-[7X21CTO1WW]-, BIOS -[TEE124N-1.40]- 06/12/2018
>>>>>> [  +0.012866] Call Trace:
>>>>>> [  +0.003945]  [<ffffffffb3313754>] dump_stack+0x19/0x1b
>>>>>> [  +0.006995]  [<ffffffffb330e91f>] dump_header+0x90/0x229
>>>>>> [  +0.007121]  [<ffffffffb2cfa982>] ? ktime_get_ts64+0x52/0xf0
>>>>>> [  +0.007451]  [<ffffffffb2d5141f>] ? delayacct_end+0x8f/0xb0
>>>>>> [  +0.007393]  [<ffffffffb2d9ac94>] oom_kill_process+0x254/0x3d0
>>>>>> [  +0.007592]  [<ffffffffb2d9a73d>] ? oom_unkillable_task+0xcd/0x120
>>>>>> [  +0.007978]  [<ffffffffb2d9a7e6>] ? find_lock_task_mm+0x56/0xc0
>>>>>> [  +0.007729]  [<ffffffffb2d9b4d6>] *out_of_memory+0x4b6/0x4f0*
>>>>>> [  +0.007358]  [<ffffffffb330f423>] __alloc_pages_slowpath+0x5d6/0x724
>>>>>> [  +0.008190]  [<ffffffffb2da18b5>] __alloc_pages_nodemask+0x405/0x420
>>>>>>
>>>>>> Thanks,
>>>>>>
>>>>>> Danyang
>>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> What most experimenters take for granted before they begin their
>>>>> experiments is infinitely more interesting than any results to which their
>>>>> experiments lead.
>>>>> -- Norbert Wiener
>>>>>
>>>>> https://www.cse.buffalo.edu/~knepley/
>>>>> <http://www.cse.buffalo.edu/~knepley/>
>>>>>
>>>>
>>>
>>> --
>>> What most experimenters take for granted before they begin their
>>> experiments is infinitely more interesting than any results to which their
>>> experiments lead.
>>> -- Norbert Wiener
>>>
>>> https://www.cse.buffalo.edu/~knepley/
>>> <http://www.cse.buffalo.edu/~knepley/>
>>>
>>
>
> --
> What most experimenters take for granted before they begin their
> experiments is infinitely more interesting than any results to which their
> experiments lead.
> -- Norbert Wiener
>
> https://www.cse.buffalo.edu/~knepley/
> <http://www.cse.buffalo.edu/~knepley/>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20191010/a6085bad/attachment-0001.html>

More information about the petsc-users mailing list