Crash in MKL 10.2 on SUSE

[AndrewC]

NOTE: MKL is being loaded from a Python extension that links against MKL.

The segment violation occurs only on SUSE ( so far), and can't be reproduced on the RHEL development machines with identical input.


GNU gdb (GDB) SUSE (7.0-0.4.16)
....
....


Program received signal SIGSEGV, Segmentation fault.
0x00007fffef2f1782 in mkl_serv_allocate () from libmkl_core.so
(gdb)
(gdb) bt
#0 0x00007fffef2f1782 in mkl_serv_allocate () from libmkl_core.so
#1 0x00007fffe8910d63 in mkl_blas_dgemm_get_bufs () from libmkl_mc3.so
#2 0x00007fffef2f7c75 in mkl_blas_dgemm_get_bufs () from libmkl_core.so
#3 0x00007fffe8938377 in mkl_blas_xdgemm_par () from libmkl_mc3.so
#4 0x00007fffef2f98df in mkl_blas_xdgemm_par () from libmkl_core.so
#5 0x00007fffe8937508 in mkl_blas_xdgemm () from libmkl_mc3.so
#6 0x00007fffef885bd9 in mkl_blas_xdgemm () from libmkl_intel_thread.so
#7 0x00007fffef993dcb in mkl_blas_dgemm () from libmkl_intel_thread.so
#8 0x00007ffff0c50f14 in dgemm_ () from libmkl_intel_lp64.so


Any ideas????

[mecej4]

In MKL 10.2 supported on your versions of SuSE and GCC?

What are the arguments (types, shapes, sizes, values) given to dgemm at the top level ?

[Gennady_F_Intel]

just for reminding, thisis the list of supported operating systems of the Intel MKL version 10.2 *:

...............

* SUSE LINUX Enterprise Server* 9, 10 (IA-32 / Intel 64 / IA-64)

..............

[AndrewC]

The problem has been resolved by building a single , 'custom' MKL DLL ,linking against that, and changing nothing else.

The loading of the chain of dependent MKL .so's seems to be root of all these problems related to Python extensions that use MKL.

[barragan_villanueva_]

Hi,

Thanks for the very imteresting info. Could you please create small testcase to reproduce the problem on our side?

[AndrewC]

Hi Victor,

Just search for "Python" in the MKL forum, you will see all the various "hoops" people, myself included, have had to go through to get a python extension that calls MKL to work.

I am afraid I cannot create a simple test case.

My understanding was that 10.3 alleviated these issues but I have not been able to test 10.3.

Andrew

[barragan_villanueva_]

Andrew,

Well, KB article http://software.intel.com/en-us/articles/using-intel-mkl-in-your-python-programs/
suggests namely creating of MKL custom library for using it from Python.
But, there exists a way to use LD_PRELOAD for MKL libraries instead.