We have huge matrix with about 100GB, and zgemm crashed. There is enough memory.
The Stack trace shows the problematic routine is : mkl_blas_avx2_zgemm_zcopy_right6_ea
We used 12 cpus with the multi-threaded lib.
What can we do to avoid such issue?
Thanks for the help.
ok, could you give us more details: OS, 64 or 32 bit?, how to link the case, ILP64 or LP64 API
Also, could you set MKL_VERBOSE and shared the output of zgemm. All this info would help us to reproduce the problem.
The fastest way would be if you will give us the reproducer.
I think it is key to link with ILP64 otherwise linear addresses (calculated internally) will exceed the 32-bit space. A square matrix of 100 GB with double-precision complex numbers would be approx. 80kx80k and a linear address that is supposed to point to the inside of such a matrix is exceeding the (signed) 32-bit space in general.