Linux 7.2 ships cache-aware scheduling and up to 5% more IOPS on EXT4 and XFS

August 16, 2026. Linus Torvalds released the stable Linux 7.2 kernel, after seven release candidates and roughly two months of development. On August 9, 2026, the seventh candidate had already promised a stable release “next weekend” — and the promise held. Kernel.org now lists Linux 7.2 as the latest mainline release, and for the first time the kernel that runs nearly every server ships cache-aware scheduling natively.

What matters here is not one more feature. It is that much of this release’s payoff lands without touching configuration, through the ordinary distro kernel update — provided you know where to look.

The cadence is itself part of the story. Linux 7.0 landed in April 2026 with the finalized EEVDF scheduler and Rust support; 7.1 followed two months later; 7.2 keeps that rhythm with a release that is unusually dense on the server side — scheduling, storage, and memory — rather than on new hardware flags alone.

Cache-aware scheduling reaches the kernel

This is the cycle’s most-discussed addition. Modern CPUs often share multiple last-level caches between groups of cores, and the classic scheduler does not always account for that topology: it can move a task between cores in a way that increases cache misses instead of reducing them.

Cache-Aware Scheduling (CAS) gives the scheduler extra information about cache topology so it can place tasks where they reuse data already resident. Testing during development produced “impressive” results on some server workloads, including major gains on certain MySQL configurations.

The caveat is stated honestly: the improvement varies widely by hardware and workload. You do not get a uniform gain everywhere. That is exactly what makes the subject serious — cache-aware scheduling is a lever, not a magic wand, and it has to be measured workload by workload.

USB4STREAM: the cable becomes a network

The second notable addition is USB4STREAM, a capability for transferring data directly between two machines over USB4 or Thunderbolt. Instead of treating USB as a simple host-to-device interface, the kernel can now use it for high-speed system-to-system communication.

The intended uses are concrete: large file transfers, development environments, system-to-system networking, debugging, and specialized workstation workflows. With USB4 and Thunderbolt increasingly common on laptops and desktops alike, this opens a path for moving large volumes between Linux machines without Ethernet — a scenario lab and post-production administrators know well.

Storage and memory: the numbers that matter

The most useful part for an SRE sits under the hood, in storage and memory management. Three optimizations are worth keeping.

• EXT4 and XFS gain roughly 5% more IOPS from a tiny optimization: simply changing the order of two operations in a very hot code path. This is the kind of gain that never shows up in flashy benchmarks but adds up across thousands of machines.
• Reading /proc/filesystems is markedly faster — an operation that turns out to happen far more often than expected in probes and monitoring tooling.
• MGLRU, the multi-generational memory reclamation infrastructure, receives an optimization that produces 30% to 100% throughput gains on some MongoDB workloads under particular configurations.

These three numbers share one property: they are automatic. No configuration change, no flag to enable. The gain materializes with the kernel update — which is precisely why it is worth measuring before you sell it to leadership.

Hardware getting ready

Linux 7.2 is also a hardware enablement release. Three advances stand out for fleets that are refreshing.

The AMD ISP4 driver finally reaches the mainline kernel: it is the image signal processor handling camera streams on recent AMD laptops. Bringing it upstream reduces the need for vendor-specific patches and simplifies life for distributions. Initial HDMI 2.1 FRL (Fixed Rate Link) support lands in AMDGPU, unlocking the bandwidth needed for 4K high-refresh-rate displays — a topic long complicated by HDMI specification restrictions. And the kernel keeps preparing for the arrival of future AMD Zen 6 processors, in the now-familiar pattern of upstreaming support before each generation ships.

On the Intel side, one notable improvement makes Intel TDX more practical by allowing certain updates without a reboot. Two deeper milestones round out the picture: the Apple M3 can now boot the mainline kernel (a foundation, not yet a polished experience — that remains the domain of projects like Asahi Linux), and RISC-V keeps maturing with ESWIN SoC support enabled by default and ongoing work on the SpacemiT K1 and K3 platforms. The Nova driver, the open-source NVIDIA GPU driver written in Rust, also keeps growing inside the DRM subsystem.

How to adopt it without breaking things

For a server fleet, adopting Linux 7.2 is not decided on kernel.org: it is decided on your distribution’s schedule. The mainline kernel is available immediately, but distributions rebuild it with their own patches, configuration options, and support policies. Ubuntu ships it through its HWE (Hardware Enablement) kernels, Red Hat and derivatives through backported kernels on stable branches, Debian through its linux-image packages.

Three steps make for a clean rollout.

uname -r                              # kernel currently running
sudo apt update && sudo apt upgrade   # Debian / Ubuntu: pull in the new kernel
sudo dnf upgrade kernel               # Fedora / RHEL: dedicated kernel package
iostat -x 1                           # disk IOPS before / after the switch

The useful reflex is to measure before you sell it: replay a representative MySQL or MongoDB benchmark, record the IOPS (iostat), and, if you want precision, the cache-misses (perf stat -e cache-misses). That is the only way to know whether cache-aware scheduling and MGLRU give you the advertised gains — or whether your workload is one that gains nothing. Because the gain is automatic, the difference only shows up in your metrics, never in a flag to enable.

One rule of caution: waiting for the first point release (7.2.1) remains sensible for homogeneous, cache-insensitive fleets. A kernel’s early-cycle regressions almost always surface in the first days of the stable branch, and a week of lead time costs nothing to teams with no MySQL or MongoDB workload to tune.

What Linux 7.2 does not do

Staying level-headed means stating what this release is not. The cache-aware scheduling and MGLRU gains are workload-specific: one MySQL or MongoDB configuration may gain a lot, another nothing at all. Apple M3 support remains initial, and Nova is not yet a replacement for Nouveau or the proprietary driver for everyday use. Finally, nothing here exempts you from reading your distribution’s release notes: distributions rebuild the kernel with their own patches and configuration choices.

Verdict

Linux 7.2 is a performance and hardware-enablement release, not a security overhaul. The decision comes down to a simple threshold: if you run MySQL, MongoDB, or EXT4/XFS storage at scale, adopt it as soon as your distribution ships it as a stable package — regression risk is low and the potential gain is measurable with zero configuration effort. If your fleet runs homogeneous, cache-insensitive workloads, nothing is urgent: wait for the first point release (7.2.1) to shake out early-cycle regressions.

The real question is no longer “should we update the kernel.” It has become “how do we measure the gain before and after.” On that front, Linux 7.2 gives administrators a rare reason to replay their internal benchmarks — and a 5% number to drop into the infrastructure report.

References

• Linux Journal, “Linux Kernel 7.2 Officially Released with Cache-Aware Scheduling, USB4STREAM, and Major AMD Improvements,” August 18, 2026.
• Kernel.org, “The Linux Kernel Archives,” mainline 7.2 release page, accessed August 20, 2026.
• OSTechNix, “Linux Kernel 7.2 RC7 Released: Stable Version Scheduled for Next Week,” August 9, 2026.