As number of threads is increased, performance per core on AESNI MB PMD decreases (it should stay constant). The root cause of this issue is a global variable in the IPSec Multi-buffer library. A patch has been generated to fix it on the library, so users running the PMD with multiple threads are encouraged to get the latest version (1.3.0) with the patch.
This issue was raised on the Intel IPSec Multi-buffer library Github repository: https://github.com/intel/intel-ipsec-mb/issues/112