parameters.markdown

Performance tuning via compilation options

Our configuration script makes educated guesses for the compiler flags to be used. More can be added via the CFLAGS keyword. A clean-sheet can be obtained by using the --disable-optimization parameter.

Some configuration options can be used to tune the code performance:

• --enable-ipo: Activate the compiler's inter-procedural optimizer
• --with-tbbmalloc: Links-in TBB's malloc library in lieu of glibc's (if found in the $PATH)
• --with-tcmalloc: Links-in tcmalloc library in lieu of glibc's (if found in the $PATH)
• --with-jemalloc: Links-in the jemalloc library in lieu of glibc's (if found in the $PATH)

The malloc libraries usually provide better performance over the system's one. SWIFT does a lot of small allocations from multiple threads.

Note that when compiling on a machine that support AVX-512 and the Intel compiler, the additional flag -qopt-zmm-usage=high is added to the list of flags for the file src/runner_doiact_grav.c which contains all the heavy math operations related to the gravity calculation. From our experience, the rest of the code does not benefit from this flag.

Performance tuning at runtime

If the code was compiled with libNUMA then the internal mechanism for thread pinning can be used to pin the threads to their cores. This is activated by adding --pin to the runtime parameters.

Note that with more recent linux kernel versions the benefit from this flag has become quite marginal.