[petsc-users] Cannot eagerly initialize cuda, as doing so results in cuda error 35 (cudaErrorInsufficientDriver) : CUDA driver version is insufficient for CUDA runtime version
Fande Kong
fdkong.jd at gmail.com
Tue Jan 25 20:20:47 CST 2022
Not sure if this is helpful. I did "git bisect", and here was the result:
[kongf at sawtooth2 petsc]$ git bisect bad
246ba74192519a5f34fb6e227d1c64364e19ce2c is the first bad commit
commit 246ba74192519a5f34fb6e227d1c64364e19ce2c
Author: Junchao Zhang <jczhang at mcs.anl.gov>
Date: Wed Oct 13 05:32:43 2021 +0000
Config: fix CUDA library and header dirs
:040000 040000 187c86055adb80f53c1d0565a8888704fec43a96
ea1efd7f594fd5e8df54170bc1bc7b00f35e4d5f M config
Started from this commit, and GPU did not work for me on our HPC
Thanks,
Fande
On Tue, Jan 25, 2022 at 7:18 PM Fande Kong <fdkong.jd at gmail.com> wrote:
>
>
> On Tue, Jan 25, 2022 at 9:04 AM Jacob Faibussowitsch <jacob.fai at gmail.com>
> 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 <fdkong.jd at gmail.com> wrote:
>>
>> Thanks Jacob,
>>
>> On Thu, Jan 20, 2022 at 6:25 PM Jacob Faibussowitsch <jacob.fai at gmail.com>
>> 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 <knepley at gmail.com> wrote:
>>>
>>> On Thu, Jan 20, 2022 at 6:44 PM Fande Kong <fdkong.jd at gmail.com> wrote:
>>>
>>>> Thanks, Jed
>>>>
>>>> On Thu, Jan 20, 2022 at 4:34 PM Jed Brown <jed at jedbrown.org> 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 <fdkong.jd at gmail.com> 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 at entry=0x2aaab7f37b70
>>>>> > <CUDADevice>, device=0x115da00, id=-35, id at entry=-1) at
>>>>> >
>>>>> /home/kongf/workhome/sawtooth/moosegpu/petsc/src/sys/objects/device/impls/cupm/cupmdevice.cxx:344
>>>>> > #2 0x00002aaab55577de in PetscDeviceCreate (type=type at entry
>>>>> =PETSC_DEVICE_CUDA,
>>>>> > devid=devid at entry=-1, device=device at 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 at entry=PETSC_DEVICE_CUDA,
>>>>> defaultDeviceId=defaultDeviceId at entry=-1)
>>>>> > at
>>>>> >
>>>>> /home/kongf/workhome/sawtooth/moosegpu/petsc/src/sys/objects/device/interface/device.cxx:273
>>>>> > #4 0x00002aaab5557bf6 in PetscDeviceInitialize
>>>>> > (type=type at 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 at entry=0x115d150,
>>>>> > method=method at 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 at entry=0x115d150,
>>>>> > method=method at 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 <fdkong.jd at gmail.com>
>>>>> wrote:
>>>>> >
>>>>> >> Thanks, Jed,
>>>>> >>
>>>>> >> This worked!
>>>>> >>
>>>>> >> Fande
>>>>> >>
>>>>> >> On Wed, Jan 19, 2022 at 11:03 PM Jed Brown <jed at jedbrown.org>
>>>>> wrote:
>>>>> >>
>>>>> >>> Fande Kong <fdkong.jd at gmail.com> writes:
>>>>> >>>
>>>>> >>> > On Wed, Jan 19, 2022 at 11:39 AM Jacob Faibussowitsch <
>>>>> >>> jacob.fai at gmail.com>
>>>>> >>> > 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/>
>>>
>>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20220125/e53abb7b/attachment-0001.html>
More information about the petsc-users
mailing list