Program aborting in DGEMM, complaining that memory cannot be allocated

[crtierney42]

I am trying to use DGEMM in MKL and my code is segfaulting. I am getting
the following error:

cannot allocate memory for thread-local data: ABORT

System configuration:

Intel ifort: 10.1.012
Intel MKL: 10.0.1.014
OS: Linux Redhat Enterprise 4.4 (upgraded to stock 2.6.20.20 kernel)
glibc: 2.3.4

I am linking with:

-L/opt/intel/mkl/10.0.1.014/lib/em64t -lguide -lpthread -lmkl

athough I don't want the parallel version. The call to dgemm is buried in
an MPI program.

The values of M,N,K are approximately 95, 1030, and 64 respectively (they vary a little between MPI tasks).

Any ideas what is causing this error message?

Thanks,
Craig

[TimP]

If you don't want the threaded MKL, link with the mkllp64, core, and sequential libraries. For the threaded version, you may need to increase thread stack size by the kmp_stacksize environment variable or library subroutine call.

[crtierney42]

Thanks for the tip. I tried linking with:

-L/opt/intel/mkl/10.0.1.014/lib/em64t -lmkl_core -lmkl_sequential -lmkl_intel_lp64

And verified that the executable was using the correct libraries. I still get the error message. My stacksize (via limit) is set to unlimited.

I found a reference on a NCSA that the problem can be alleviated by statically linking the libraries. So I am trying to link the libraries statically. I am trying to link libmkl_sequential.a,libmkl_intel_lp64.a, and libmkl_core.a to my executable. However,
the linker complains that dgemm_ is not found. I have tried moving libmkl_core.a to the front of the list, but this didn't help.

What is the proper order for linking?

Thanks,
Craig

[TimP]

There are examples of static link in the notes in /doc/ showing how to use -Wl,--begin-group .... -Wl,--end-group. This depends on your compiler and binutils being sufficiently up to date. It didn't work consistently with Intel compilers until 10.1.012. Otherwise, you have to list libmkl_intel_lp64.a libmkl_core.a libmkl_sequential.a (full path names) 3 times to resolve circular dependencies, more times if out of order. I think of lp_64 or ilp64 as the one with the interfaces you call in your code, and mkl_sequential as the one which stubs out OpenMP library calls, as a way to remember the order.