Performance slow down when dynamically linking with Intel MKL

When dynamically linking with Intel MKL, users may find performance slow down in a few test applications. This is because loading dynamic libraries at runtime brings overhead to the application. The problem only happens when data set in the application is very small, and the overhead of shared library load cannot be ignored.

The following conditions can help to identify this problem: 1) the data set is small in the application. 2) The second run may have better performance than the first run. 3) The problem does not happen when statically linking with Intel MKL.

Note, when users benchmark small data with test application, the shared library load may impact performance result. Users can link MKL statically to avoid the problem. In real applications, overhead of shared library load is negligible.

This article applies to: Intel® Math Kernel Library Knowledge Base