Clear Linux & Their Love For FMV + dl_platform/dl_hwcap In The Name Of Performance
For those mesmerized by the numbers whenever posting a cross-distribution comparison like the recent Core i9 7900X vs. Threadripper 1950X On Ubuntu 17.10, Antergos, Clear Linux with showing Intel's performance optimizations done on Clear Linux, Intel engineer Victor Rodriguez presented this week at the 2017 Open-Source Summit North America about some of their Linux performance boosting work.
While Clear Linux ships with aggressive compiler flags and other optimizations, contrary to the belief of some, their distribution does work on Intel hardware going back to ~2011 and just not the most recent generations of CPUs. But for remaining optimized for both new and old hardware, they do utilize Function Multi-Versioning (FMV) as offered by GCC. They also rely upon optimized binaries for particular hardware platforms via GLIBC with the dl_platform/dl_hwcap features for shipping optimized libraries that are then selected at run-time based on the CPU. This is one of the approaches to how Clear Linux is already shipping with AVX-512 optimized libraries.
Victor's talk mostly focused on the GCC/GLIBC paths for shipping optimized code. I haven't seen any audio/video recordings of the presentation, but there is the PDF slide deck for those interested. More details on their FMV, AutoFDO, and other optimization work is described at ClearLinux.org.
While Clear Linux ships with aggressive compiler flags and other optimizations, contrary to the belief of some, their distribution does work on Intel hardware going back to ~2011 and just not the most recent generations of CPUs. But for remaining optimized for both new and old hardware, they do utilize Function Multi-Versioning (FMV) as offered by GCC. They also rely upon optimized binaries for particular hardware platforms via GLIBC with the dl_platform/dl_hwcap features for shipping optimized libraries that are then selected at run-time based on the CPU. This is one of the approaches to how Clear Linux is already shipping with AVX-512 optimized libraries.
Victor's talk mostly focused on the GCC/GLIBC paths for shipping optimized code. I haven't seen any audio/video recordings of the presentation, but there is the PDF slide deck for those interested. More details on their FMV, AutoFDO, and other optimization work is described at ClearLinux.org.
19 Comments