[Dune] UG grid code: Segfault on large core counts
Oliver Sander
sander at igpm.rwth-aachen.de
Wed Dec 5 11:20:34 CET 2012
Am 05.12.2012 10:48, schrieb Peter Bastian:
> Hi Oliver,
>
> you ask questions… The parallel UG graphics used some clever
> way to do depth ordering by sorting the coarse mesh in
> view direction and then "interpolating" this ordering in the
> grid hierarchy. I do not know whether it has something to
> do with that. Obviously some tree is constructed on
> the processor graph and the maximum degree is estimated
> using the strange formula. I really cannot say more as it was
> all Klaus Johannsens work.
Hi Peter,
okay, that explains why the file is 20k lines long (you guys were tough!).
I then propose to make WOP_DOWN_CHANNELS_MAX very large(tm),
and forget about it. Since the number of channels computed in line 21023
is roughly the square root of the processor number I guess a value
of about 2000 should be okay for the nearer future.
best,
Oliver
>
> Best,
>
> Peter
>
> Am 05.12.2012 um 09:46 schrieb Oliver Sander<sander at igpm.rwth-aachen.de>:
>
>> Am 04.12.2012 20:10, schrieb Peter Bastian:
>>> Dear Eike,
>>>
>>> a quick search through the code suggests that
>>> the constant
>>> #define WOP_DOWN_CHANNELS_MAX 32
>>> in ug/graphics/uggraph/wop.h line 146 is too small.
>>>
>>> In wop.c line 21023
>>>
>>> WopDownChannels = (INT)ceil(0.5*(sqrt(4.0*(DOUBLE)procs-3.0)-1.0));
>>>
>>> evaluates to 39 which fairly close above 32. Then out of bound access in many
>>> arrays is produced. Maybe it is as simple as that.
>> Hi Peter,
>> I came to the same conclusion. Can you give us an explanation of what
>> these WopDownChannels are?
>> best,
>> Oliver
>>
>>> Best,
>>>
>>> Peter
>>>
>>>
>>> Am 04.12.2012 um 18:39 schrieb Oliver Sander<sander at igpm.rwth-aachen.de>:
>>>
>>>> Am 04.12.2012 17:05, schrieb Eike Mueller:
>>>>> Dear dune-list,
>>>> Hi Eike,
>>>>> now my code runs on 6, 24, 96 and 384 processors. However, on 1536 cores it crashes with a (different) segmentation faullt,
>>>> is that with custom types for DDD_GID, DDD_PROC and such, or with the defaults?
>>>>
>>>> Is it possible to get a valgrind output? The crash is at program startup, so even a
>>>> slow valgrind run shouldn't take too long.
>>>>
>>>> The crash is in the ug graphics implementation, which we don't use from Dune.
>>>> So in the very worst case we could out-comment lots of stuff and make it work
>>>> that way. But I prefer a proper fix.
>>>>
>>>> best,
>>>> Oliver
>>>>
>>>>> when I inspect the core dump with
>>>>>
>>>>> gdb<executable> --core=<core>
>>>>>
>>>>> I get this backtrace:
>>>>>
>>>>> Core was generated by `/work/n02/n02/eike/Code/DUNEGeometricMultigrid/geometricmultigrid_nz128_r130_DE'.
>>>>> Program terminated with signal 11, Segmentation fault.
>>>>> #0 0x0000000000898d7a in PPIF::RecvASync (v=0x100000001, data=0x13b80a0,
>>>>> size=4, error=0x7fffffff41f0) at ppif.c:656
>>>>> 656 ((MPIVChannel*)v)->p, ((MPIVChannel*)v)->chanid, COMM, req) )
>>>>> (gdb) backtrace
>>>>> #0 0x0000000000898d7a in PPIF::RecvASync (v=0x100000001, data=0x13b80a0,
>>>>> size=4, error=0x7fffffff41f0) at ppif.c:656
>>>>> #1 0x000000000081ce2b in NumberOfDesc () at wop.c:21095
>>>>> #2 0x000000000082b2db in UG::D2::InitWOP () at wop.c:24677
>>>>> #3 0x00000000007fbb1d in UG::D2::InitUGGraph () at initgraph.c:90
>>>>> #4 0x00000000007fbac5 in UG::D2::InitGraphics () at graphics.c:133
>>>>> #5 0x0000000000715e4c in UG::D2::InitUg (argcp=0x7fffffff465c,
>>>>> argvp=0x7fffffff4648) at ../initug.c:293
>>>>> #6 0x000000000051bb8c in Dune::UG_NS<2>::InitUg (argcp=0x7fffffff465c,
>>>>> argvp=0x7fffffff4648) at ../../../dune/grid/uggrid/ugwrapper.hh:910
>>>>> #7 0x000000000052062a in Dune::UGGrid<3>::UGGrid (this=0x1603360)
>>>>> at uggrid.cc:74
>>>>> #8 0x000000000059332f in Dune::GridFactory<Dune::UGGrid<3> >::GridFactory (
>>>>> this=0x7fffffff9cd8) at uggridfactory.cc:74
>>>>> #9 0x000000000040faa2 in SphericalGridGenerator::SphericalGridGenerator (
>>>>> this=0x7fffffff9cd0, filename=..., refcount=5)
>>>>> at sphericalgridgenerator.hh:37
>>>>> #10 0x00000000004010c8 in main (argc=2, argv=0x7fffffffb778)
>>>>> at geometricmultigrid.cc:167
>>>>>
>>>>> I'm setting up a grid with one element on each core, and then refine it 5 times, so the total grid size is 49152. I don't know whether this has any influence, but I set the default heap size to 8000.
>>>>>
>>>>> Any ideas what might be going on here, or just suggestions on how to proceed with debugging a code with this large number of cores would be much appreciated. I'm now trying to get a better trace with the Cray ATP tool.
>>>>>
>>>>> Thank you very much,
>>>>>
>>>>> Eike
>>>>>
>>>>> On 30 Nov 2012, at 18:47, Eike Mueller wrote:
>>>>>
>>>>>> Hi Oliver,
>>>>>>
>>>>>> I just did some detective work and I think I know why it crashes and how it can be fixed.
>>>>>>
>>>>>> The problem is that in parallel/ddd/include/ddd.h the type DDD_PROC is defined as 'unsigned short', so can only store up to 2^16 different processor IDs. In the subroutine IFCreateFromScratch(), where the segfault occurs, the variable lastproc, which is initialised to PROC_INVALID is of this type. According to parallel/ddd/dddi.h, PROC_INVALID=(MAX_PROCS+1) and MAX_PROCS=(1<<MAX_PROCBITS_IN_GID). Investigating the core dump with gdb, lastproc is indeed 1, at least for MAX_PROCBITS_IN_GID=16 and then on processor 1 the pointer ifHead never gets initialised, so dereferencing it with ifHead->nItems++; causes the segfault.
>>>>>>
>>>>>> I now changed DDD_PROC to 'unsigned int' in parallel/ddd/include/ddd.h and then the test program I sent you works fine if I run on 6 processors (I can also look at the .vtu files if I write them out in ascii format).
>>>>>> I can run with up to 6 refinement steps, and if I go to 7 it runs out of memory with
>>>>>>
>>>>>> testsphericalgridgenerator: xfer/xfer.c:942: void UG::D3::ddd_XferRegisterDelete(UG::D3::DDD_HDR): Assertion `0' failed.
>>>>>> DDD [000] FATAL 06060: out of memory during XferEnd()
>>>>>> _pmiu_daemon(SIGCHLD): [NID 01842] [c11-0c1s6n0] [Fri Nov 30 18:35:27 2012] PE RANK 0 exit signal Aborted
>>>>>> DDD [001] WARNING 02514: increased coupling table, now 131072 entries
>>>>>> DDD [001] WARNING 02224: increased object table, now 131072 entries
>>>>>> [NID 01842] 2012-11-30 18:35:27 Apid 3144197: initiated application termination
>>>>>> Application 3144197 exit codes: 134
>>>>>> Application 3144197 resources: utime ~24s, stime ~1s
>>>>>>
>>>>>> but I think this is another problem, and in my case 6 should actually be enough.
>>>>>>
>>>>>> Do you think the change I made makes sense and doesn't break anything else? If that is the case, what's the best way of setting up a patch? To be on the safe side, maybe something like this:
>>>>>>
>>>>>> #if (DDD_MAX_PROCBITS_IN_GID< 16)
>>>>>> typedef unsigned short DDD_PROC;
>>>>>> #else
>>>>>> typedef unsigned int DDD_PROC;
>>>>>> #endif // (DDD_MAX_PROCBITS_IN_GID< 16)
>>>>>>
>>>>>> I need to do some more testing to check that it all works for my full code, and I will keep you updated on any progress.
>>>>>>
>>>>>> Thanks,
>>>>>>
>>>>>> Eike
>>>>>>
>>>>>> Oliver Sander wrote:
>>>>>>> Okay, with maxprocbits = 16 I do see a crash. 2 processors and 1 level is enough.
>>>>>>> I'm in a bit of a hurry right now, but I'll have a look at it later.
>>>>>>> best,
>>>>>>> Oliver
>>>>>>> Am 29.11.2012 19:36, schrieb Eike Mueller:
>>>>>>>> Hi Oliver,
>>>>>>>>
>>>>>>>> thank you very much for trying this out and for the patch. I now rebuilt UG with the latest patch file you sent me this morning and I still get the segfault. However, this only occurs if I specify the with_ddd_maxprocbits flag, if I do not set this (i.e. only use DDD_GID=long, as you do), then it runs fine (I have only done a small run so far, and haven't tried 7 refinement steps, so can not say anything about that other error you get).
>>>>>>>> I tried both with_ddd_maxprocbits=20 and 16, but it does not work in any of these cases. The default is 2^9=512, and unfortunately that's not enough for me, I would need at least 2^16=65536.
>>>>>>>> As for the error message you get when reading the .vtu files, could that be because they are written out in Dune::VTK::appendedbase64 format? I also get an error message when I open them in paraview
>>>>>>>>
>>>>>>>> ERROR: In /home/kitware/ParaView3/Utilities/BuildScripts/ParaView-3.6/ParaView3/VTK/IO/vtkXMLUnstructuredDataReader.cxx, line 522
>>>>>>>> vtkXMLUnstructuredGridReader (0xa16c918): Cannot read points array from Points in piece 0. The data array in the element may be too short.
>>>>>>>>
>>>>>>>> When I opened .vtu files produced by my main solver code on HECToR, paraview actually crashed, and this was before I modified any of the UG settings. I could fix this by writing data out in Dune::VTK::ascii format instead. Could this be a big/little endian issue? HECToR is 64bit, but my local desktop, where I run paraview to look at the output is 32bit, not sure if that has any impact.
>>>>>>>>
>>>>>>>> Eike
>>>>>>>>
>>>>>>>> Oliver Sander wrote:
>>>>>>>>> Hi Eike,
>>>>>>>>> I tried your example with DDD_GID==long, on my laptop where sizeof(long)==8 and sizeof(uint)==4.
>>>>>>>>> I started the program with
>>>>>>>>>
>>>>>>>>> mpirun -np 6 ./testsphericalgridgenerator sphericalshell_cube_6.dgf 6
>>>>>>>>>
>>>>>>>>> Besides a few DDD warnings that I have never seen before, it works like a charm.
>>>>>>>>>
>>>>>>>>> What version of UG are you using? I'll send you the very latest patch file just
>>>>>>>>> to be sure.
>>>>>>>>>
>>>>>>>>> The programs runs, but paraview gives me an error message when trying to open
>>>>>>>>> the output file. Does that happen to you, too?
>>>>>>>>>
>>>>>>>>> I didn't try the with_ddd_maxprocbits setting. Does your program crash if you
>>>>>>>>> do _not_ set this?
>>>>>>>>>
>>>>>>>>> For the worst case: is it possible to get a temporary account on your hector computer?
>>>>>>>>>
>>>>>>>>> best,
>>>>>>>>> Oliver
>>>>>>>>>
>>>>>>>>> Am 28.11.2012 10:28, schrieb Eike Mueller:
>>>>>>>>>> Hi Oliver,
>>>>>>>>>>
>>>>>>>>>> thanks a lot, that would be great. My desktop is a 32bit machine, where sizeof(long) = sizeof(int) = 4, so I'm not sure if recompiling everything with GID=long there will make a difference.
>>>>>>>>>>
>>>>>>>>>> Eike
>>>>>>>>>>
>>>>>>>>>> Oliver Sander wrote:
>>>>>>>>>>> Thanks for the backtrace. I'll try and see whether I can reproduce the crash
>>>>>>>>>>> on my machine. If that's not possible things will be a bit difficult :-)
>>>>>>>>>>> --
>>>>>>>>>>> Oliver
>>>>>>>>>>>
>>>>>>>>>>> Am 26.11.2012 18:39, schrieb Eike Mueller:
>>>>>>>>>>>> Hi Markus and Oliver,
>>>>>>>>>>>>
>>>>>>>>>>>> to get to the bottom of this I recompiled everything (UG+Dune+my code) with -O0 -g, and that way I was able to get some more information out of the core dump. On 1 processor it runs fine now, but when running on 6, this is what I get, looks like it crashes in loadBalance(), but I can't make sense of what's happening inside UG. It always seems to crash inside ifcreate.c, either in line 482 or 489:
>>>>>>>>>>>>
>>>>>>>>>>>> Program terminated with signal 11, Segmentation fault.
>>>>>>>>>>>> #0 0x0000000000a438b0 in UG::D3::IFCreateFromScratch (tmpcpl=0x1462ad0,
>>>>>>>>>>>> ifId=1) at if/ifcreate.c:489
>>>>>>>>>>>> 489 ifHead->nItems++;
>>>>>>>>>>>> (gdb) backtrace
>>>>>>>>>>>> #0 0x0000000000a438b0 in UG::D3::IFCreateFromScratch (tmpcpl=0x1462ad0,
>>>>>>>>>>>> ifId=1) at if/ifcreate.c:489
>>>>>>>>>>>> #1 0x0000000000a44dae in UG::D3::IFRebuildAll () at if/ifcreate.c:1059
>>>>>>>>>>>> #2 0x0000000000a44e71 in UG::D3::IFAllFromScratch () at if/ifcreate.c:1097
>>>>>>>>>>>> #3 0x0000000000a4bd89 in UG::D3::DDD_XferEnd () at xfer/cmds.c:869
>>>>>>>>>>>> #4 0x0000000000a65b5c in UG::D3::TransferGridFromLevel (theMG=0x1441880,
>>>>>>>>>>>> level=0) at trans.c:835
>>>>>>>>>>>> #5 0x0000000000a5df4b in UG::D3::lbs (argv=0x7fffffffa390 "0",
>>>>>>>>>>>> theMG=0x1441880) at lb.c:659
>>>>>>>>>>>> #6 0x0000000000a0bd83 in UG::D3::LBCommand (argc=4, argv=0x7fffffffab90)
>>>>>>>>>>>> at commands.c:10658
>>>>>>>>>>>> #7 0x00000000004a1e2d in Dune::UG_NS<3>::LBCommand (argc=4,
>>>>>>>>>>>> argv=0x7fffffffab90) at ../../../dune/grid/uggrid/ugwrapper.hh:979
>>>>>>>>>>>> #8 0x00000000004a8df9 in Dune::UGGrid<3>::loadBalance (this=0x134e490,
>>>>>>>>>>>> strategy=0, minlevel=0, depth=2, maxLevel=32, minelement=1)
>>>>>>>>>>>> at uggrid.cc:556
>>>>>>>>>>>> #9 0x00000000004077bf in Dune::UGGrid<3>::loadBalance (this=0x134e490)
>>>>>>>>>>>> at /home/n02/n02/eike/work/Library/Dune2.2/include/dune/grid/uggrid.hh:738
>>>>>>>>>>>> #10 0x0000000000400928 in main (argc=3, argv=0x7fffffffb5a8)
>>>>>>>>>>>> at testsphericalgridgenerator.cc:65
>>>>>>>>>>>>
>>>>>>>>>>>> but I sometimes also get:
>>>>>>>>>>>>
>>>>>>>>>>>> Core was generated by `./testsphericalgridgenerator sphericalshell_cube_6.dgf 4'.
>>>>>>>>>>>> Program terminated with signal 11, Segmentation fault.
>>>>>>>>>>>> #0 0x0000000000a438b0 in UG::D3::IFCreateFromScratch (tmpcpl=0x1462ad0,
>>>>>>>>>>>> ifId=1) at if/ifcreate.c:482
>>>>>>>>>>>> 482 ifAttr->nAB = ifAttr->nBA = ifAttr->nABA = 0;
>>>>>>>>>>>> (gdb) backtrace
>>>>>>>>>>>> #0 0x0000000000a438b0 in UG::D3::IFCreateFromScratch (tmpcpl=0x1462ad0,
>>>>>>>>>>>> ifId=1) at if/ifcreate.c:482
>>>>>>>>>>>> #1 0x0000000000a44dae in UG::D3::IFRebuildAll () at if/ifcreate.c:1049
>>>>>>>>>>>> #2 0x0000000000a44e71 in UG::D3::IFRebuildAll () at if/ifcreate.c:1057
>>>>>>>>>>>> #3 0x0000000000a4bd89 in UG::D3::DDD_XferEnd () at xfer/cmds.c:850
>>>>>>>>>>>> #4 0x0000000000a65b5c in UG::D3::TransferGridFromLevel (theMG=0x1441880,
>>>>>>>>>>>> level=0) at trans.c:824
>>>>>>>>>>>> #5 0x0000000000a5df4b in UG::D3::lbs (argv=0x7fffffffa390 "0",
>>>>>>>>>>>> theMG=0x1441880) at lb.c:644
>>>>>>>>>>>> #6 0x0000000000a0bd83 in UG::D3::LBCommand (argc=4, argv=0x7fffffffab90)
>>>>>>>>>>>> at commands.c:10644
>>>>>>>>>>>> #7 0x00000000004a1e2d in Dune::UG_NS<3>::LBCommand (argc=0,
>>>>>>>>>>>> argv=0x7fffffffa420) at ../../../dune/grid/uggrid/ugwrapper.hh:977
>>>>>>>>>>>> #8 0x00000000004a8df9 in Dune::UGGrid<3>::loadBalance (this=0x134e490,
>>>>>>>>>>>> strategy=0, minlevel=0, depth=2, maxLevel=32, minelement=1)
>>>>>>>>>>>> at uggrid.cc:554
>>>>>>>>>>>> #9 0x00000000004077bf in std::__shared_count<(__gnu_cxx::_Lock_policy)2>::~__shared_count (this=0x134e490, __in_chrg=<optimized out>)
>>>>>>>>>>>> at /opt/gcc/4.6.3/snos/include/g++/bits/shared_ptr_base.h:550
>>>>>>>>>>>> #10 0x0000000000400928 in main (argc=0, argv=0x520)
>>>>>>>>>>>> at testsphericalgridgenerator.cc:66
>>>>>>>>>>>>
>>>>>>>>>>>> I've tried different load balancing strategies, but for all I get a segfault.
>>>>>>>>>>>>
>>>>>>>>>>>> Cheers,
>>>>>>>>>>>>
>>>>>>>>>>>> Eike
>>>>>>>>>>>>
>>>>>>>>>>>>
>>>>>>>>>>>> Markus Blatt wrote:
>>>>>>>>>>>>> On Mon, Nov 26, 2012 at 01:57:15PM +0000, Eike Mueller wrote:
>>>>>>>>>>>>>> thanks a lot for the patch, unfortunately I still get a segfault when I run on HECToR.
>>>>>>>>>>>>>>
>>>>>>>>>>>>> I feared that, but it was still worth a shot. The change probably
>>>>>>>>>>>>> interferes with the memory allocation in ddd.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Markus
>>>>>> --
>>>>>> Dr Eike Mueller
>>>>>> Research Officer
>>>>>>
>>>>>> Department of Mathematical Sciences
>>>>>> University of Bath
>>>>>> Bath BA2 7AY, United Kingdom
>>>>>>
>>>>>> +44 1225 38 5633
>>>>>> e.mueller at bath.ac.uk
>>>>>> http://people.bath.ac.uk/em459/
>>>>>>
>>>>>> _______________________________________________
>>>>>> Dune mailing list
>>>>>> Dune at dune-project.org
>>>>>> http://lists.dune-project.org/mailman/listinfo/dune
>>>> _______________________________________________
>>>> Dune mailing list
>>>> Dune at dune-project.org
>>>> http://lists.dune-project.org/mailman/listinfo/dune
>>> ------------------------------------------------------------
>>> Peter Bastian
>>> Interdisziplinäres Zentrum für Wissenschaftliches Rechnen
>>> Universität Heidelberg
>>> Im Neuenheimer Feld 368
>>> D-69120 Heidelberg
>>> Tel: 0049 (0) 6221 548261
>>> Fax: 0049 (0) 6221 548884
>>> email: peter.bastian at iwr.uni-heidelberg.de
>>> web: http://conan.iwr.uni-heidelberg.de/people/peter/
>>>
>>>
>>> _______________________________________________
>>> Dune mailing list
>>> Dune at dune-project.org
>>> http://lists.dune-project.org/mailman/listinfo/dune
>>
>> _______________________________________________
>> Dune mailing list
>> Dune at dune-project.org
>> http://lists.dune-project.org/mailman/listinfo/dune
> ------------------------------------------------------------
> Peter Bastian
> Interdisziplinäres Zentrum für Wissenschaftliches Rechnen
> Universität Heidelberg
> Im Neuenheimer Feld 368
> D-69120 Heidelberg
> Tel: 0049 (0) 6221 548261
> Fax: 0049 (0) 6221 548884
> email: peter.bastian at iwr.uni-heidelberg.de
> web: http://conan.iwr.uni-heidelberg.de/people/peter/
>
More information about the Dune
mailing list