RomanNess / performanceGap

An example for the strange performance gap between SLES and CentOS

performanceGap

A minimal example extracted from my libsampling driver, that illustrates a performance gap of up to 2x between CentOS and SLES (on our compute nodes with equivalent hardware). The additional overhead in SLES is proportional to the number of instrumented function calls (= dynamic calls to __cyg_profile_func_XXX).

The behavior only occurs if all of the following factors are present

• A dynamic library with __cyg_profile - functions
• An global array accessed via thread local variable

Build

$ make

Note that this does also create a soft link to the "parallel" version of the library (were the performance gap happens).

Run

$ make run

... to run the latest build behind the soft link. Or just choose a target with "parallel" flavor and run it yourself.

Requirements

• GCC (I used 4.9.3 on both machines)