The Performance Cost To A Proposed Fedora 37 CFLAGS/CXXFLAGS Change

Redis is another real-world workload that was hit particularly hard by this compiler flag.

In total I ran 100 different tests on this mid-range TUXEDO laptop with AMD Ryzen 5 5500U laptop for seeing the impact of building software with "-fno-omit-frame-pointer" on Fedora 36 with GCC 12. While in some cases there was minimal impact to no measurable difference, in plenty of workloads this did lead to a significant performance regression. Below is a look at the subset of the tests with a measurable performance improvement when rebuilding the software with "-O2" rather than "-O2 -fno-omit-frame-pointer":

The Botan crypto library was hit most hard along with Redis, Liquid-DSP digital signal processing library, PJSIP VoIP communications, FLAC audio encoding, ASTC texture compression, Intel oneDNN deep neural network library, and various other software packages were up to a few percent slower.

Of the 100 tests carried out for this article, when taking the geometric mean of all these benchmarks it equated to about a 14% performance penalty of the software with -O2 compared to when adding -fno-omit-frame-pointer. Those wanting to go through all 100 benchmark results in full can do so via this result page.

It will be very interesting to see what comes of this Fedora 37 change proposal. While this change can enhance the performance profiling and debugging of Fedora Linux out-of-the-box, it does come with non-zero performance costs... I wonder what's the ratio of those using "perf" on a daily basis and other profiling/debugging tools on Fedora compared to those as a general purpose OS? In any event see this Fedora Wiki page for those wanting to learn more about this proposed feature change.