Ceci est une ancienne révision du document !
MKL vs LAPACK
We strongly encourage the use of MKL instead of Lapack because MKL is more efficient and optimized for the CPU architecture.
The default BLAS is not parallel. It uses only 1 core
Benchmarks
Let's use dgesv to solve a random linear system.
- randomsys.f90
program randomsys implicit none real(kind=8), dimension(:),allocatable :: x,b real(kind=8), dimension(:,:),allocatable :: a real(kind=8) :: err integer :: i, info, lda, ldb, nrhs, n integer, dimension(:), allocatable :: ipiv ! initialize random number generator seed ! if you remove this, the same numbers will be generated each ! time you run this code. call init_random_seed() print *, "Input n ... " read *, n allocate(a(n,n)) allocate(b(n)) allocate(x(n)) allocate(ipiv(n)) call random_number(a) call random_number(x) b = matmul(a,x) ! compute RHS nrhs = 1 ! number of right hand sides in b lda = n ! leading dimension of a ldb = n ! leading dimension of b call dgesv(n, nrhs, a, lda, ipiv, b, ldb, info) ! Note: the solution is returned in b ! and a has been changed. ! compare computed solution to original x: print *, " x computed rel. error" i=n ! do i=1,n err = abs(x(i)-b(i))/abs(x(i)) print '(3d16.6)', x(i),b(i),err ! enddo deallocate(a,b,ipiv) end program randomsys
Compile with LAPACK (and BLAS)
$ module load gcc $ gfortran -O2 randomsys.f90 -o randomsys_lapack.exe -lblas -llapack
Compile with Intel MKL
$ module load intel/14.0.2 $ ifort -O2 -mkl randomsys.f90 -o randomsys_mkl.exe
Results
| N | LAPACK | MKL (1 core) | MKL (16 core) |
| 4000 | 14s | 6s | 2s |
| 6000 | 41s | 12s | 4s |
| 8000 | 1m27s | 25s | 5s |
| 10000 | 45s | 8s | |
For MKL, we just used
OMP_NUM_THREADS env variable to select the numer of threads to use