Intel Appears On The Verge Of Some Exciting Performance Optimizations For Linux Distros

One thing that has never gotten old over the past nearly twenty years of covering Linux news on Phoronix are the relentless performance optimizations made to the Linux kernel, GCC and LLVM/Clang compilers, and other key open-source projects over the years. Intel engineers have been responsible for so many exciting Linux performance optimizations over time at ensuring maximum Linux x86_64 performance as well as ensuring great performance at a macro-level as they've showcased with the likes of Clear Linux. It looks like they have some new innovation(s) in store soon for further maximizing compiler-assisted performance.

Over the years on Phoronix you've likely read an article or two about BOLT that was developed by Facebook/Meta for optimizing the layout of generated binaries. BOLT made it to the mainline LLVM codebase last year with LLVM 14 for helping to optimize binaries. Facebook has even demonstrated its potential for speeding up the Linux kernel via BOLT as well as enhancing the performance of user-space applications and other software.

The downside of this post-link optimizer is that it relies on first having to collect an application's execution profile via the Linux perf sampling. It's similar to compiler Profile Guided Optimizations (PGO) in that you first need to collect an accurate profile of the software being built before feeding that into BOLT for being able to optimize the binary of large applications.

It turns out Intel's Clear Linux engineers have been investigating BOLT and ultimately coming up with some new yet-to-be-announced solution for carrying out similar optimizations that can be done at a Linux OS/distribution level scale. A Phoronix reader pointed out this Clear Linux ticket inquiring about BOLT support, to which Intel Fellow Arjan van de Ven commented:

"We have been looking at bolt for... quite some time, and even did a series of prototypes of things similar or prototype new optimizations inside bolt.

Bolt has some logistical issues to use it well ---- but what makes it slightly messy for us is that it still at times creates invalid output.

BUT -- we will be doing something bolt-like in the very near future (we're finishing up final pieces of it right now) that, while not bolt level, should get close, but with the logistics solved for a distro and without the risk of invalid output...

We want to get this widely deployed in the OS still this year :)"

Color me interested. That comment was made two weeks ago.

It will be interesting to see how Intel's Clear Linux team ends out 2023 and there will certainly be benchmarks once the new optimization(s) are rolled out. If the work pans out hopefully we'll see other Linux distributions adopting similar approaches for optimizing their binaries too.