Using High-bandwidth Memory with Intel MKL

To achieve maximum performance, Intel MKL may use the memkind library (https://github.com/memkind/memkind), which enables controlling memory characteristics and partitioning the heap between different kinds of memory. By default Intel MKL memory manager tries to allocate memory to Multi-Channel Dynamic Random Access Memory (MCDRAM) using the memkind library on the 2^nd generation Intel® Xeon Phi™ product family (for more details of MCDRAM, see https://software.intel.com/en-us/articles/mcdram-high-bandwidth-memory-on-knights-landing-analysis-methods-tools). If allocation of memory to MCDRAM is not possible at the moment, Intel MKL memory manager falls back to a regular system allocator.

By default the amount of MCDRAM available for Intel MKL is unlimited. To control the amount of MCDRAM available for Intel MKL, do either of the following:

• Call

  mkl_set_memory_limit (MKL_MEM_MCDRAM, <limit_in_mbytes>)

• Set the environment variable:

  • For the bash shell:

    MKL_FAST_MEMORY_LIMIT="<limit_in_mbytes>"

  • For a C shell (csh or tcsh):

    setenv MKL_FAST_MEMORY_LIMIT "<limit_in_mbytes>"

The setting of the limit affects all Intel MKL functions, including user-callable memory functions such as mkl_malloc. Therefore, if an application calls mkl_malloc, mkl_calloc, or mkl_realloc, which always tries to allocate memory to MCDRAM, make sure that the limit is sufficient.

If you replace Intel MKL memory management functions with your own functions (for details, see Redefining Memory Functions), Intel MKL uses your functions and does not work with the memkind library directly.