On Tue, Jan 25, 2022 at 9:04 AM Jacob Faibussowitsch <[email protected]> wrote:
> Configure should not have an impact here I think. The reason I had you run > `cudaGetDeviceCount()` is because this is the CUDA call (and in fact the > only CUDA call) in the initialization sequence that returns the error code. > There should be no prior CUDA calls. Maybe this is a problem with > oversubscribing GPU’s? In the runs that crash, how many ranks are using any > given GPU at once? Maybe MPS is required. > I used one MPI rank. Fande > > Best regards, > > Jacob Faibussowitsch > (Jacob Fai - booss - oh - vitch) > > On Jan 21, 2022, at 12:01, Fande Kong <[email protected]> wrote: > > Thanks Jacob, > > On Thu, Jan 20, 2022 at 6:25 PM Jacob Faibussowitsch <[email protected]> > wrote: > >> Segfault is caused by the following check at >> src/sys/objects/device/impls/cupm/cupmdevice.cxx:349 being a >> PetscUnlikelyDebug() rather than just PetscUnlikely(): >> >> ``` >> if (PetscUnlikelyDebug(_defaultDevice < 0)) { // _defaultDevice is in >> fact < 0 here and uncaught >> ``` >> >> To clarify: >> >> “lazy” initialization is not that lazy after all, it still does some 50% >> of the initialization that “eager” initialization does. It stops short >> initializing the CUDA runtime, checking CUDA aware MPI, gathering device >> data, and initializing cublas and friends. Lazy also importantly swallows >> any errors that crop up during initialization, storing the resulting error >> code for later (specifically _defaultDevice = -init_error_value;). >> >> So whether you initialize lazily or eagerly makes no difference here, as >> _defaultDevice will always contain -35. >> >> The bigger question is why cudaGetDeviceCount() is returning >> cudaErrorInsufficientDriver. Can you compile and run >> >> ``` >> #include <cuda_runtime.h> >> >> int main() >> { >> int ndev; >> return cudaGetDeviceCount(&ndev): >> } >> ``` >> >> Then show the value of "echo $?”? >> > > Modify your code a little to get more information. > > #include <cuda_runtime.h> > #include <cstdio> > > int main() > { > int ndev; > int error = cudaGetDeviceCount(&ndev); > printf("ndev %d \n", ndev); > printf("error %d \n", error); > return 0; > } > > Results: > > $ ./a.out > ndev 4 > error 0 > > > I have not read the PETSc cuda initialization code yet. If I need to guess > at what was happening. I will naively think that PETSc did not get correct > GPU information in the configuration because the compiler node does not > have GPUs, and there was no way to get any GPU device information. > > During the runtime on GPU nodes, PETSc might have incorrect information > grabbed during configuration and had this kind of false error message. > > Thanks, > > Fande > > > >> >> Best regards, >> >> Jacob Faibussowitsch >> (Jacob Fai - booss - oh - vitch) >> >> On Jan 20, 2022, at 17:47, Matthew Knepley <[email protected]> wrote: >> >> On Thu, Jan 20, 2022 at 6:44 PM Fande Kong <[email protected]> wrote: >> >>> Thanks, Jed >>> >>> On Thu, Jan 20, 2022 at 4:34 PM Jed Brown <[email protected]> wrote: >>> >>>> You can't create CUDA or Kokkos Vecs if you're running on a node >>>> without a GPU. >>> >>> >>> I am running the code on compute nodes that do have GPUs. >>> >> >> If you are actually running on GPUs, why would you need lazy >> initialization? It would not break with GPUs present. >> >> Matt >> >> >>> With PETSc-3.16.1, I got good speedup by running GAMG on GPUs. That >>> might be a bug of PETSc-main. >>> >>> Thanks, >>> >>> Fande >>> >>> >>> >>> KSPSetUp 13 1.0 6.4400e-01 1.0 2.02e+09 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 0 5 0 0 0 0 5 0 0 0 3140 64630 15 1.05e+02 5 >>> 3.49e+01 100 >>> KSPSolve 1 1.0 1.0109e+00 1.0 3.49e+10 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 0 87 0 0 0 0 87 0 0 0 34522 69556 4 4.35e-03 1 >>> 2.38e-03 100 >>> KSPGMRESOrthog 142 1.0 1.2674e-01 1.0 1.06e+10 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 0 27 0 0 0 0 27 0 0 0 83755 87801 0 0.00e+00 0 >>> 0.00e+00 100 >>> SNESSolve 1 1.0 4.4402e+01 1.0 4.00e+10 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 21100 0 0 0 21100 0 0 0 901 51365 57 1.10e+03 52 >>> 8.78e+02 100 >>> SNESSetUp 1 1.0 3.9101e-05 1.0 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 0 0 0.00e+00 0 >>> 0.00e+00 0 >>> SNESFunctionEval 2 1.0 1.7097e+01 1.0 1.60e+07 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 8 0 0 0 0 8 0 0 0 0 1 0 0 0.00e+00 6 >>> 1.92e+02 0 >>> SNESJacobianEval 1 1.0 1.6213e+01 1.0 2.80e+07 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 8 0 0 0 0 8 0 0 0 0 2 0 0 0.00e+00 1 >>> 3.20e+01 0 >>> SNESLineSearch 1 1.0 8.5582e+00 1.0 1.24e+08 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 4 0 0 0 0 4 0 0 0 0 14 64153 1 3.20e+01 3 >>> 9.61e+01 94 >>> PCGAMGGraph_AGG 5 1.0 3.0509e+00 1.0 8.19e+07 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 1 0 0 0 0 1 0 0 0 0 27 0 5 3.49e+01 9 >>> 7.43e+01 0 >>> PCGAMGCoarse_AGG 5 1.0 3.8711e+00 1.0 0.00e+00 0.0 0.0e+00 0.0e+00 >>> 0.0e+00 2 0 0 0 0 2 0 0 0 0 0 0 0 0.00e+00 0 >>> 0.00e+00 0 >>> PCGAMGProl_AGG 5 1.0 7.0748e-01 1.0 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 0 0 0.00e+00 0 >>> 0.00e+00 0 >>> PCGAMGPOpt_AGG 5 1.0 1.2904e+00 1.0 2.14e+09 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 1 5 0 0 0 1 5 0 0 0 1661 29807 26 7.15e+02 20 >>> 2.90e+02 99 >>> GAMG: createProl 5 1.0 8.9489e+00 1.0 2.22e+09 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 4 6 0 0 0 4 6 0 0 0 249 29666 31 7.50e+02 29 >>> 3.64e+02 96 >>> Graph 10 1.0 3.0478e+00 1.0 8.19e+07 1.0 0.0e+00 >>> 0.0e+00 0.0e+00 1 0 0 0 0 1 0 0 0 0 27 0 5 >>> 3.49e+01 9 7.43e+01 0 >>> MIS/Agg 5 1.0 4.1290e-01 1.0 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 0 0 >>> 0.00e+00 0 0.00e+00 0 >>> SA: col data 5 1.0 1.9127e-02 1.0 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 0 0 >>> 0.00e+00 0 0.00e+00 0 >>> SA: frmProl0 5 1.0 6.2662e-01 1.0 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 0 0 >>> 0.00e+00 0 0.00e+00 0 >>> SA: smooth 5 1.0 4.9595e-01 1.0 1.21e+08 1.0 0.0e+00 >>> 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 244 2709 15 >>> 1.97e+02 15 2.55e+02 90 >>> GAMG: partLevel 5 1.0 4.7330e-01 1.0 6.98e+08 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 0 2 0 0 0 0 2 0 0 0 1475 4120 5 1.78e+02 10 >>> 2.55e+02 100 >>> PCGAMG Squ l00 1 1.0 2.6027e+00 1.0 0.00e+00 0.0 0.0e+00 0.0e+00 >>> 0.0e+00 1 0 0 0 0 1 0 0 0 0 0 0 0 0.00e+00 0 >>> 0.00e+00 0 >>> PCGAMG Gal l00 1 1.0 3.8406e-01 1.0 5.48e+08 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 0 1 0 0 0 0 1 0 0 0 1426 4270 1 1.48e+02 2 >>> 2.11e+02 100 >>> PCGAMG Opt l00 1 1.0 2.4932e-01 1.0 7.20e+07 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 0 0 0 0 0 0 0 0 0 0 289 2653 1 6.41e+01 1 >>> 1.13e+02 100 >>> PCGAMG Gal l01 1 1.0 6.6279e-02 1.0 1.09e+08 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 0 0 0 0 0 0 0 0 0 0 1645 3851 1 2.40e+01 2 >>> 3.64e+01 100 >>> PCGAMG Opt l01 1 1.0 2.9544e-02 1.0 7.15e+06 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 0 0 0 0 0 0 0 0 0 0 242 1671 1 4.84e+00 1 >>> 1.23e+01 100 >>> PCGAMG Gal l02 1 1.0 1.8874e-02 1.0 3.72e+07 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 0 0 0 0 0 0 0 0 0 0 1974 3636 1 5.04e+00 2 >>> 6.58e+00 100 >>> PCGAMG Opt l02 1 1.0 7.4353e-03 1.0 2.40e+06 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 0 0 0 0 0 0 0 0 0 0 323 1457 1 7.71e-01 1 >>> 2.30e+00 100 >>> PCGAMG Gal l03 1 1.0 2.8479e-03 1.0 4.10e+06 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 0 0 0 0 0 0 0 0 0 0 1440 2266 1 4.44e-01 2 >>> 5.51e-01 100 >>> PCGAMG Opt l03 1 1.0 8.2684e-04 1.0 2.80e+05 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 0 0 0 0 0 0 0 0 0 0 339 1667 1 6.72e-02 1 >>> 2.03e-01 100 >>> PCGAMG Gal l04 1 1.0 1.2238e-03 1.0 2.09e+05 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 0 0 0 0 0 0 0 0 0 0 170 244 1 2.05e-02 2 >>> 2.53e-02 100 >>> PCGAMG Opt l04 1 1.0 4.1008e-04 1.0 1.77e+04 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 0 0 0 0 0 0 0 0 0 0 43 165 1 4.49e-03 1 >>> 1.19e-02 100 >>> PCSetUp 2 1.0 9.9632e+00 1.0 4.95e+09 1.0 0.0e+00 0.0e+00 >>> 0.0e+00 5 12 0 0 0 5 12 0 0 0 496 17826 55 1.03e+03 45 >>> 6.54e+02 98 >>> PCSetUpOnBlocks 44 1.0 9.9087e-04 1.0 2.88e+03 1.0 >>> >>> >>> >>> >>> >>> >>>> The point of lazy initialization is to make it possible to run a solve >>>> that doesn't use a GPU in PETSC_ARCH that supports GPUs, regardless of >>>> whether a GPU is actually present. >>>> >>>> Fande Kong <[email protected]> writes: >>>> >>>> > I spoke too soon. It seems that we have trouble creating cuda/kokkos >>>> vecs >>>> > now. Got Segmentation fault. >>>> > >>>> > Thanks, >>>> > >>>> > Fande >>>> > >>>> > Program received signal SIGSEGV, Segmentation fault. >>>> > 0x00002aaab5558b11 in >>>> > >>>> Petsc::CUPMDevice<(Petsc::CUPMDeviceType)0>::CUPMDeviceInternal::initialize >>>> > (this=0x1) at >>>> > >>>> /home/kongf/workhome/sawtooth/moosegpu/petsc/src/sys/objects/device/impls/cupm/cupmdevice.cxx:54 >>>> > 54 PetscErrorCode CUPMDevice<T>::CUPMDeviceInternal::initialize() >>>> noexcept >>>> > Missing separate debuginfos, use: debuginfo-install >>>> > bzip2-libs-1.0.6-13.el7.x86_64 elfutils-libelf-0.176-5.el7.x86_64 >>>> > elfutils-libs-0.176-5.el7.x86_64 glibc-2.17-325.el7_9.x86_64 >>>> > libX11-1.6.7-4.el7_9.x86_64 libXau-1.0.8-2.1.el7.x86_64 >>>> > libattr-2.4.46-13.el7.x86_64 libcap-2.22-11.el7.x86_64 >>>> > libibmad-5.4.0.MLNX20190423.1d917ae-0.1.49224.x86_64 >>>> > libibumad-43.1.1.MLNX20200211.078947f-0.1.49224.x86_64 >>>> > libibverbs-41mlnx1-OFED.4.9.0.0.7.49224.x86_64 >>>> > libmlx4-41mlnx1-OFED.4.7.3.0.3.49224.x86_64 >>>> > libmlx5-41mlnx1-OFED.4.9.0.1.2.49224.x86_64 libnl3-3.2.28-4.el7.x86_64 >>>> > librdmacm-41mlnx1-OFED.4.7.3.0.6.49224.x86_64 >>>> > librxe-41mlnx1-OFED.4.4.2.4.6.49224.x86_64 libxcb-1.13-1.el7.x86_64 >>>> > libxml2-2.9.1-6.el7_9.6.x86_64 numactl-libs-2.0.12-5.el7.x86_64 >>>> > systemd-libs-219-78.el7_9.3.x86_64 xz-libs-5.2.2-1.el7.x86_64 >>>> > zlib-1.2.7-19.el7_9.x86_64 >>>> > (gdb) bt >>>> > #0 0x00002aaab5558b11 in >>>> > >>>> Petsc::CUPMDevice<(Petsc::CUPMDeviceType)0>::CUPMDeviceInternal::initialize >>>> > (this=0x1) at >>>> > >>>> /home/kongf/workhome/sawtooth/moosegpu/petsc/src/sys/objects/device/impls/cupm/cupmdevice.cxx:54 >>>> > #1 0x00002aaab5558db7 in >>>> > Petsc::CUPMDevice<(Petsc::CUPMDeviceType)0>::getDevice >>>> > (this=this@entry=0x2aaab7f37b70 >>>> > <CUDADevice>, device=0x115da00, id=-35, id@entry=-1) at >>>> > >>>> /home/kongf/workhome/sawtooth/moosegpu/petsc/src/sys/objects/device/impls/cupm/cupmdevice.cxx:344 >>>> > #2 0x00002aaab55577de in PetscDeviceCreate (type=type@entry >>>> =PETSC_DEVICE_CUDA, >>>> > devid=devid@entry=-1, device=device@entry=0x2aaab7f37b48 >>>> > <defaultDevices+8>) at >>>> > >>>> /home/kongf/workhome/sawtooth/moosegpu/petsc/src/sys/objects/device/interface/device.cxx:107 >>>> > #3 0x00002aaab5557b3a in PetscDeviceInitializeDefaultDevice_Internal >>>> > (type=type@entry=PETSC_DEVICE_CUDA, >>>> defaultDeviceId=defaultDeviceId@entry=-1) >>>> > at >>>> > >>>> /home/kongf/workhome/sawtooth/moosegpu/petsc/src/sys/objects/device/interface/device.cxx:273 >>>> > #4 0x00002aaab5557bf6 in PetscDeviceInitialize >>>> > (type=type@entry=PETSC_DEVICE_CUDA) >>>> > at >>>> > >>>> /home/kongf/workhome/sawtooth/moosegpu/petsc/src/sys/objects/device/interface/device.cxx:234 >>>> > #5 0x00002aaab5661fcd in VecCreate_SeqCUDA (V=0x115d150) at >>>> > >>>> /home/kongf/workhome/sawtooth/moosegpu/petsc/src/vec/vec/impls/seq/seqcuda/veccuda.c:244 >>>> > #6 0x00002aaab5649b40 in VecSetType (vec=vec@entry=0x115d150, >>>> > method=method@entry=0x2aaab70b45b8 "seqcuda") at >>>> > >>>> /home/kongf/workhome/sawtooth/moosegpu/petsc/src/vec/vec/interface/vecreg.c:93 >>>> > #7 0x00002aaab579c33f in VecCreate_CUDA (v=0x115d150) at >>>> > >>>> /home/kongf/workhome/sawtooth/moosegpu/petsc/src/vec/vec/impls/mpi/mpicuda/ >>>> > mpicuda.cu:214 >>>> > #8 0x00002aaab5649b40 in VecSetType (vec=vec@entry=0x115d150, >>>> > method=method@entry=0x7fffffff9260 "cuda") at >>>> > >>>> /home/kongf/workhome/sawtooth/moosegpu/petsc/src/vec/vec/interface/vecreg.c:93 >>>> > #9 0x00002aaab5648bf1 in VecSetTypeFromOptions_Private >>>> (vec=0x115d150, >>>> > PetscOptionsObject=0x7fffffff9210) at >>>> > >>>> /home/kongf/workhome/sawtooth/moosegpu/petsc/src/vec/vec/interface/vector.c:1263 >>>> > #10 VecSetFromOptions (vec=0x115d150) at >>>> > >>>> /home/kongf/workhome/sawtooth/moosegpu/petsc/src/vec/vec/interface/vector.c:1297 >>>> > #11 0x00002aaab02ef227 in libMesh::PetscVector<double>::init >>>> > (this=0x11cd1a0, n=441, n_local=441, fast=false, >>>> ptype=libMesh::PARALLEL) >>>> > at >>>> > >>>> /home/kongf/workhome/sawtooth/moosegpu/scripts/../libmesh/installed/include/libmesh/petsc_vector.h:693 >>>> > >>>> > On Thu, Jan 20, 2022 at 1:09 PM Fande Kong <[email protected]> >>>> wrote: >>>> > >>>> >> Thanks, Jed, >>>> >> >>>> >> This worked! >>>> >> >>>> >> Fande >>>> >> >>>> >> On Wed, Jan 19, 2022 at 11:03 PM Jed Brown <[email protected]> wrote: >>>> >> >>>> >>> Fande Kong <[email protected]> writes: >>>> >>> >>>> >>> > On Wed, Jan 19, 2022 at 11:39 AM Jacob Faibussowitsch < >>>> >>> [email protected]> >>>> >>> > wrote: >>>> >>> > >>>> >>> >> Are you running on login nodes or compute nodes (I can’t seem to >>>> tell >>>> >>> from >>>> >>> >> the configure.log)? >>>> >>> >> >>>> >>> > >>>> >>> > I was compiling codes on login nodes, and running codes on compute >>>> >>> nodes. >>>> >>> > Login nodes do not have GPUs, but compute nodes do have GPUs. >>>> >>> > >>>> >>> > Just to be clear, the same thing (code, machine) with PETSc-3.16.1 >>>> >>> worked >>>> >>> > perfectly. I have this trouble with PETSc-main. >>>> >>> >>>> >>> I assume you can >>>> >>> >>>> >>> export PETSC_OPTIONS='-device_enable lazy' >>>> >>> >>>> >>> and it'll work. >>>> >>> >>>> >>> I think this should be the default. The main complaint is that >>>> timing the >>>> >>> first GPU-using event isn't accurate if it includes initialization, >>>> but I >>>> >>> think this is mostly hypothetical because you can't trust any >>>> timing that >>>> >>> doesn't preload in some form and the first GPU-using event will >>>> almost >>>> >>> always be something uninteresting so I think it will rarely lead to >>>> >>> confusion. Meanwhile, eager initialization is viscerally disruptive >>>> for >>>> >>> lots of people. >>>> >>> >>>> >> >>>> >>> >> >> -- >> 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/> >> >> >
