openblas segfaults · Issue #218 · xianyi/OpenBLAS

相关文章推荐

活泼的手术刀 · 不能将类型“AxiosRequestHead ...· 1 年前 ·

有情有义的大象 · BAT批处理判断服务是否正常运行的方法(批处 ...· 3 年前 ·

单身的斑马 · R语言中的quantile()函数_r ...· 3 年前 ·

飘逸的荔枝 · Systemd ...· 3 年前 ·

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement . We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

I am getting segfaults when running an openmp-parallelised code that calls a blas matrix-multiplication from the parallel region. The machine is an 8-core "Intel(R) Xeon(R) CPU E5345", with 16 GB RAM, running Ubuntu 12.04 LTS. I have essentially the same problem with the 0.1 version shipped with 12.04 and the backported 0.2.6 version. I used gfortran 4.6.3 to compile and link. The stack size is set to unlimited, and OMP_STACKSIZE=1G. The problem on this machine occurs if OMP_NUM_THREADS=5 or greater.

Here is the example program:

program test_blas_error
   implicit none
   integer, parameter :: dp = kind(1.0d0)
   real(dp), dimension(:,:), allocatable :: a, b, c
   integer :: i, n, n_max, l_max
   n = 20
   n_max = 1000
   l_max = 100
!$omp parallel do schedule(dynamic) default(none) shared(n, n_max, l_max) private(i, a, b, c)
   do i = 1, n
      allocate(a(n_max,l_max), b(n_max,l_max), c(n_max,n_max))
      call random_number(a)
      call random_number(b)
      call dgemm('N','T',n_max, n_max, l_max, 1.0_dp, &
      a, size(a,1), b, size(b,1), 0.0_dp, c, size(c, 1))
      print*, sum(c)
      deallocate(a,b,c)
   enddo
!$omp end parallel do
endprogram test_blas_error

$ gfortran -fopenmp test_blas_error.f95 -lblas
$ ./a.out
BLAS : Program is Terminated. Because you tried to allocate too many memory regions.
   24944796.711494826
BLAS : Program is Terminated. Because you tried to allocate too many memory regions.
Segmentation fault
$ gfortran -g -fbacktrace -fopenmp test_blas_error.f95 -lblas
ab686@womble:~/test$ ./a.out
BLAS : Program is Terminated. Because you tried to allocate too many memory regions.
BLAS : Program is Terminated. Because you tried to allocate too many memory regions.
Program received signal SIGSEGV: Segmentation fault - invalid memory reference.
Backtrace for this error:
   24944796.711494826
BLAS : Program is Terminated. Because you tried to allocate too many memory regions.
Program received signal SIGSEGV: Segmentation fault - invalid memory reference.
Backtrace for this error:
BLAS : Program is Terminated. Because you tried to allocate too many memory regions.
#0  0x2B10EF77A547
#1  0x2B10EF77AB4E
#2  0x2B10EFEDC49F
Program received signal SIGSEGV: Segmentation fault - invalid memory reference.
Backtrace for this error:
#0  0x2B10EF77A547
#1  0x2B10EF77AB4E
#2  0x2B10EFEDC49F
Program received signal SIGSEGV: Segmentation fault - invalid memory reference.
Backtrace for this error:
#0  0x2B10EF77A547
#1  0x2B10EF77AB4E
#2  0x2B10EFEDC49F
#3  0x57C4EF in dgemm_otcopy_CORE2 at gemm_tcopy_4.S:247
#0  0x2B10EF77A547
#1  0x2B10EF77AB4E
#2  0x2B10EFEDC49F
#3  0x57C4EF in dgemm_otcopy_CORE2 at gemm_tcopy_4.S:247
   24974178.514966901
#3  0x57C4EF in dgemm_otcopy_CORE2 at gemm_tcopy_4.S:247
#3  0x57C4EF in dgemm_otcopy_CORE2 at gemm_tcopy_4.S:247
Segmentation fault
I'd like to add that in Makefile.conf NUM_CORES=8 (auto-detected) and USE_OPENMP=0 ( blas shouldn't use openmp here I think.)
I'd be grateful for any comments, advice.

Albert
          Hi Xianyi,
Many thanks for your quick reply. I have tried 2) but got the same:
$ export OPENBLAS_NUM_THREADS=1
$ ./a.out
BLAS : Program is Terminated. Because you tried to allocate too many memory regions.
   24944796.711494826
BLAS : Program is Terminated. Because you tried to allocate too many memory regions.
Program received signal SIGSEGV: Segmentation fault - invalid memory reference.
Backtrace for this error:
BLAS : Program is Terminated. Because you tried to allocate too many memory regions.
BLAS : Program is Terminated. Because you tried to allocate too many memory regions.
Program received signal SIGSEGV: Segmentation fault - invalid memory reference.
Backtrace for this error:
Program received signal SIGSEGV: Segmentation fault - invalid memory reference.
Backtrace for this error:
#0  0x2AF00EFDB547
#1  0x2AF00EFDBB4E
#2  0x2AF00F73D49F
Program received signal SIGSEGV: Segmentation fault - invalid memory reference.
Backtrace for this error:
#0  0x2AF00EFDB547
#1  0x2AF00EFDBB4E
#2  0x2AF00F73D49F
#0  0x2AF00EFDB547
#1  0x2AF00EFDBB4E
#2  0x2AF00F73D49F
   24942308.924021363
#3  0x57C4EF in dgemm_otcopy_CORE2 at gemm_tcopy_4.S:247
#0  0x2AF00EFDB547
#1  0x2AF00EFDBB4E
#2  0x2AF00F73D49F
#3  0x57C4EF in dgemm_otcopy_CORE2 at gemm_tcopy_4.S:247
#3  0x57C4EF in dgemm_otcopy_CORE2 at gemm_tcopy_4.S:247
   25022022.333436657
#3  0x57C4EF in dgemm_otcopy_CORE2 at gemm_tcopy_4.S:247
Segmentation fault
I am currently compiling with USE_OPENMP=1 and I'll let you know the result when it's finished.
Albert
          OK, compiling with USE_OPENMP=1 was a success, I didn't get any more segfaults. I am wondering in this case is the blas routine running in on a single core or parallel (called from a parallel region). Is this controlled by OPENBLAS_NUM_THREADS? What is the default?
Thanks for your help.
Albert
          Hi @albapa ,
By default, the number of threads is controlled by OPENBLAS_NUM_THREADS.  When it built with USE_OPENMP=1, it is controlled by OMP_NUM_THREADS.
The default value of OPENBLAS_NUM_THREADS is the number of cpu cores.
Xianyi
          Hi Xianyi,

Thanks for your reply.
One more thing and sorry if this should be obvious - assume openblas is built with USE_OPENMP=1, and I call a blas routine from a parallel region - so the blas routine is called OMP_NUM_THREADS times concurrently - does each blas instance run on a single (the calling) thread, or is it threaded further? Thank you very much in advance.
Albert
          If you call BLAS routine from a parallel region, BLAS routine will not run multi-threaded further.

If you call BLAS routine from a sequential region, BLAS routine is multi-threaded.