Hi, I am experiencing the problem that I mention in the title. I have a program that we have built and linked in three different way: (i) ICC + MKL, (ii) GCC + MKL + GOMP and (iii) GCC + Atlas. Out of them, only (i) and (iii) seem to work, because option (ii) breaks down with a segmentation fault in mkl_serv_malloc() after exhausting all memory.
Hi Gennady, I am not sure whether this is useful for you. I bundled our simulation code here
It consists on a general purpose high-level algebra library for C++ (mps-bundle-gcc/) and the simulation itself (mps-bundle-gcc/xy_...). To build it and run it is a three steps process:
* From mps-bundle-gcc/ issue "CC=gcc make build" this should detect the Intel MKL library and build the software with it
* From mps-bundle-gcc/xy_stuck issue "make"
* Execute mps-bundle-gcc/xy_stuck/xy_stuck.exe
In our case memory consumption grows until the program issues a SIGSEGV
I attach here the valgrind summary