For this we designed a tool called MiniApex, that is intended to extract the small part of a user’s large scientific code, that is representative for the system load generated by running the whole application. For this MiniApex makes use of the Clang compiler, to parse and analyse the sourcecode, to find the routines of the application that make up the kernel. After this identification has taken place, the code is then extracted into a new mini-app, that can then be equipped with a custom entry point to function as a benchmark. Given the entry function into the computational kernel, we recursively descent through all callees, marking them as necessary to extract. To do this across multiple sourcecode files, also called translation units, we use a whole program callgraph for this step. After collecting all functions, we step through this subset of functions a second time, this time marking all modifications to globally visible state, aswell as uses of structs and definitions as necessary to extract. This concludes the sourcecode extraction part. As Clang however operates on the expanded sourcecode, but our mini-app should represent the sourcecode before the preprocessor is run, we also hook into the preprocessor, the extract the unexpanded form of all makros, pragmas and definitions. These are then spliced into the extracted sourcecode parts. The last remaining step is to manually write a setup routine to call this extracted mini-app.