Cloud & InfraArticle Cache-aware scheduling and leaner memory reclaim target EPYC and Xeon fleets, but the biggest wins are opt-in.
Ji-ho Choi Linus Torvalds shipped Linux 7.2 on schedule on August 16, grumbling that the final week was "once again" busier than he'd like — a "new normal" he attributes partly to the patch and bug-report churn of the AI-tooling era. The headlines lead with faster I/O and fresh AMD and Intel drivers, and sure, AMDGPU picked up HDMI 2.1 FRL support and Intel's Xe driver keeps maturing. But if you run servers for a living, the driver notes are the least interesting part of this release. Linux 7.2 is the kernel finally catching up to what CPUs have looked like for the last five years — and most of its biggest wins won't reach you automatically.
The scheduler learns what a chiplet is #
The marquee feature is cache-aware load balancing, an Intel-led effort more than a year in the making. The idea is simple to state and was hard to land: keep tasks that share data on the same last-level-cache domain, instead of letting the load balancer scatter them across CCDs or dies and turn every shared access into a cross-cache round trip.
This matters because the hardware moved and the scheduler didn't. An AMD EPYC part isn't one big pool of cores with one L3 — it's a grid of chiplets, each with its own slice of cache, and Intel's Xeon 6 has gone modular too. For years the practical fix lived in userspace: pin your database threads with taskset
, carve out CCDs with cpusets, tune NUMA policy by hand, and re-do it all when the hardware refresh lands. CFS — and EEVDF after it — balanced for load, not locality, and on multi-LLC parts that's frequently the wrong objective. Phoronix's testing during the cycle showed solid gains on Zen 5 for PostgreSQL, Valkey, and network-heavy workloads, exactly the class of software where threads hammer shared state.
Here's the catch: it ships behind CONFIG_SCHED_CACHE
, a build-time Kconfig option, not a runtime toggle that's on by default. That's the kernel community being appropriately paranoid — scheduler regressions are the third rail of kernel development — but it means the feature only exists for you if your distro's kernel team enables it, or you build your own. Whether Ubuntu 26.10, which will ship 7.2, flips it on is now one of the more consequential packaging decisions of the fall.
Memory management for people who actually run out of memory #
The quieter, arguably bigger story is in mm. The MGLRU reclaim path got a cleanup of its reclaim loop and dirty-writeback handling that KernelNewbies pegs at up to ~30% throughput improvement on workloads like MongoDB under YCSB. That's a reclaim-path fix, not a benchmark stunt — it pays off precisely when your working set exceeds RAM, which describes most cost-optimized cloud fleets on any given Tuesday.
The swap subsystem's "swap table" rework hit phase IV, unifying allocation and charging and cutting static metadata to nearly zero — about 512MB saved per 1TB of swap device mounted. That number sounds niche until you remember the industry is deliberately re-embracing swap: zswap tiers, memory overcommit on Kubernetes nodes, and CXL-backed far memory all treat swap as a first-class tier rather than an emergency brake. Leaner swap metadata and a faster reclaim loop are infrastructure for that shift, not housekeeping. (A further RFC that replaces the global swap allocation lock with per-CPU structures is still working its way upstream — that one hasn't landed yet.)
Filesystems follow the same theme: Btrfs enables large folios by default with experimental 2MB huge-folio support, ext4 reworked fast commit to cut lock contention, and file-backed transparent huge pages now work for writable mappings. Every one of these is about doing fewer, bigger operations per unit of I/O — the through-line of this whole release.
The parts that will bite you #
7.2 also swings the axe. The deprecated strncpy()
API is finally gone from the kernel after a cleanup effort that dragged on for roughly six years, architecture-specific MD5 implementations were dropped, and the minimum LLVM/Clang version for Clang-built kernels went up. If you maintain out-of-tree modules — vendor BSPs, storage drivers, that GPU shim nobody wants to own — none of this waits for your sprint planning. Build against 7.2 now and boot it on real hardware; a strncpy
call that compiled fine against 6.x is now a hard error. Reports also point to the beginning of deprecation for AF_ALG, the userspace crypto socket, which would eventually affect anyone doing kernel-offloaded crypto from userspace — worth watching, though the timeline there is longer.
And there's a cautionary tale embedded in the release: the new "fair" DRM GPU scheduler, a CFS-inspired replacement for FIFO that was supposed to improve interactivity under GPU load, caused enough regressions that the default was reverted to FIFO late in the cycle. The code is still there; the default isn't. Scheduler heuristics that benchmark well and behave badly in the wild are a genre, and it's a fair preview of why cache-aware scheduling is gated behind a Kconfig rather than switched on for everyone.
The verdict #
This is the most consequential kernel of the 7.x series so far for server operators, and it's not close. But treat the performance wins as an engineering project, not a free upgrade. The reclaim and folio improvements arrive with the kernel; the headline scheduler feature needs a distro (or you) to opt in; and the deprecations demand a test cycle before your fleet touches it. One more planning note: if history holds, the year's LTS pick will be a later release — likely 7.3 or 7.4 in the winter — so enterprise distros will consume this work then. Benchmark 7.2 on your own workloads now, on real chiplet hardware, so you know which flags to fight for when your LTS kernel lands.
Sources & further reading #
[The 7.2 kernel has been released](https://lwn.net/Articles/1088991/)— lwn.net -
[Linux 7.2 Released With Faster I/O, New AMD & Intel Driver Improvements](https://www.phoronix.com/news/Linux-7.2-Released)— phoronix.com -
[Cache Aware Scheduling Merged For Linux 7.2 For Boosting Modern Intel & AMD CPUs](https://www.phoronix.com/news/Linux-7.2-Scheduler)— phoronix.com -
[Linux 7.2 changelog](https://kernelnewbies.org/Linux_7.2)— kernelnewbies.org -
[Kernel & Embedded News: Linux 7.2 Nears Release; Swap I/O Rework](https://dev.to/raghu_bharadwaj_404e60eb0c/kernel-embedded-news-linux-72-nears-release-swap-io-rework-52e0)— dev.to
[Ji-ho Choi](https://sourcefeed.dev/u/jiho_choi)· Security & Cloud Editor
Ji-ho covers the increasingly tangled overlap between cloud architecture and security, drawing on a background as a penetration tester to keep his reporting grounded in real-world attack paths. He never lets a vendor claim go unquestioned and insists that every buzzword come with a proof of concept.
Discussion 0 #
No comments yet
Be the first to weigh in.