[Dune] UG grid code: Segfault on large core counts
Oliver Sander
sander at igpm.rwth-aachen.de
Wed Dec 5 09:46:39 CET 2012
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
More information about the Dune
mailing list