Liquorix ships experimental server memory research into a desktop kernel, all in one afternoon
Liquorix maintainer Steven Barrett folded Kairui Song's still-in-review frequency-guided memory-management patch set into an enthusiast gaming kernel on Friday, and it's one of the more interesting packaging calls in a while.
Two builds landed within about an hour of each other. First came 7.2-19, a straightforward merge of the freshly released Linux 7.2.11 stable tree. Then, roughly 60 minutes later, 7.2-20 showed up carrying that same base with the MGLRU Frequency Guided v4 (MGLRU-FG) patch set layered on top.
The second build is the story. It's the first time Kairui Song's experimental page-reclaim technology, originally built for data-center servers and Android handsets, has reached a desktop-oriented build through a normal package install.
Two builds, one difference
Both were tagged Friday, October 16, 2026. Barrett's changelog for the first was as terse as stable-tracking gets:
- merge v7.2.11
- update version to 7.2.11-1
The next one said something different entirely:
- merge the MGLRU Frequency Guided v4 patch set
- update version to 7.2.11-2
Reading the commit diff between the two tags, the relationship is unambiguous. 7.2-20 is a strict superset of 7.2-19. The only real change is a Zen patch bump from v7.2.11-lqx1 to v7.2.11-lqx2, a swap that accounts for about 2,684 additions and 54 deletions. That's the MGLRU-FG code plus whatever minor Zen tweaks ride along with it.
The scheduler, tick rate, preemption model, I/O schedulers and network stack didn't move at all between the two. The single differentiator is memory reclaim.
For context, this isn't a random detour. The previous release, 7.2-18, had introduced zswap and zram optimizations, signaling an ongoing interest in making memory pressure less painful. The next logical step was to improve how the kernel decides what to keep in RAM and what to evict.
What MGLRU-FG actually does
RAM is finite, so when it gets tight the kernel's reclaimer has to pick which pages to evict. It uses the LRU, the Least Recently Used list, keeping pages that were touched recently and dropping those that haven't been seen in a while.
The current default, MGLRU, was introduced in Linux 6.30. It organizes pages into generations and uses an approximate aging model to find cold pages cheaply. It's lighter than the old classical LRU, but it mostly reacts to when a page was last accessed, not how often it gets used.
Kairui Song's Frequency Guided patch teaches MGLRU to also track access frequency. A page gets promoted to a hotter generation once it's been accessed enough times, roughly eight accesses per bump. A page referenced at least twice is treated as part of the working set, which makes the modern scheme behave more like the classical model for the subsystems that care about that, like PSI pressure signals and readahead.
It also frees up one bit in the folio page-flag word. Page flags are an extremely tight resource inside the kernel's per-page data structure, so even a single bit matters.
Barrett's call to ship this on a desktop kernel is a bet that better working-set tracking will smooth out interactive desktop behavior, keeping hot application data resident and dodging stutters when the system is juggling cache, RAM and swap. Whether that transfer holds up on real gaming or audio workloads is exactly the question the enthusiast community will spend the next few weeks answering.
The benchmarks in the patch cover letters are compelling, though you should keep in mind they're author-reported and run against server, database and Android-style loads. Representative figures cited include:
| Workload | Result with MGLRU-FG |
|---|---|
| MongoDB YCSB | +10.3% throughput |
| Chromium & Node.js | +47.4% requests |
| LevelDB GET/SCAN | +41% |
| SQLite hot lookup vs. cold stream | ~95% faster |
The author wraps it up as "10% to 40% higher performance or lower refault in various different tests," with a few workloads seeing dramatic gains and one page-flag bit freed along the way. Treat those as directional, not marketing.
Why it arrived the way it did
MGLRU-FG lives at the bleeding edge of kernel memory management and has historically only reached users through custom compiles or mainline, once it's finally accepted. Kairui Song, a kernel engineer at Tencent, posted it in RFC form starting in May 2026, and this v4 went out via Patchew on October 15 or 16, the same window Barrett used for the packaging.
Releasing the clean base and then the experimental variant within the same work session shows how lightweight modern distro-kernel toolchains have become. One person can ship a stable merge and a test build in the same afternoon.
It's also worth remembering that Liquorix has an explicit philosophy here, called Zen Interactive Tuning and shared with the Zen kernel. It favors responsiveness at the cost of sustained throughput and power efficiency. As Liquorix's own docs put it:
"It's a common myth that custom zen kernels magically give you 20% more speed. They don't."
So if the memory patch doesn't help on a gaming box, the trade-off the kernel already makes is a real one. Battery life and idle efficiency take a hit, and heavy compile jobs or large file transfers can run slightly longer. Barrett has said his quality assurance is "basic dogfooding on most/all of my devices," so this isn't a formal test matrix.
That's why 7.2-19 exists as an option. It's the same base, minus the RFC-stage memory code. If reproducibility and stability matter more than experimentation, it's the safer pick. And if you're running servers or KVM virtualization, the maintainer recommends switching back to your distro's stock kernel altogether.
How to try it
The install script is the usual quick route:
curl -s 'https://liquorix.net/install-liquorix.sh' | sudo bash
Debian, Ubuntu, and Arch Linux are supported.
A couple of housekeeping notes before you upgrade. Back up your system and know how to boot an older kernel, since this is an experimental patch sitting in a production build. And keep in mind that Liquorix tracks upstream stable, so you still get the same security and bug fixes as everyone else, including the October 15 Linux 7.2.11 release Greg Kroah-Hartman shipped the day before, which carried an unusually blunt "all users of the 7.2 kernel series must upgrade" directive.
