All written by Chinese developers! The Linux 7.3 kernel optimizations are incredibly aggressive: ByteDance cuts latency by 90%, ZTE boosts speed 490x, and Xiaomi makes the Raspberry Pi fly.

Linux 7.2 was officially released on August 16, and the merge window for 7.3 subsequently opened. Based on the patches merged in recent days, the core memory and scheduling optimizations in this merge window almost entirely come from contributions by Chinese engineers.

In particular, optimization patches from ByteDance, ZTE, and Xiaomi were merged into the mainline simultaneously, covering three key modules: scheduling latency, memory management, and compressed memory.

ByteDance engineer Chuyi Zhou discovered that the kernel's smp_call_function*() disables preemption throughout the entire wait for other CPUs to complete IPIs. On a 16-core machine, IPI waiting can push scheduling latency up to 17ms. For DPDK, a network framework with extremely high real-time requirements, this constitutes a fatal tail latency.

The ByteDance team's fix is elegant: recording the "CPUs to wait for" into each task's CPU mask, turning the wait information into task-local data, thereby re-enabling preemption before waiting. In testing, DPDK worst-case latency dropped from approximately 17ms to about 1.5ms, a reduction of roughly 90%.

ZTE engineer Xu Xin discovered a severe performance issue in KSM's reverse mapping function rmap_walk_ksm(): the worst-case hold time of an internal lock can reach 705ms. Applications would freeze for hundreds of milliseconds out of nowhere, causing containers to time out directly. The triggering scenario is very common: when JVM allocates heap via mmap and repeatedly calls mprotect for GC barriers, it triggers the issue.

After the ZTE team's optimization, worst-case lock hold time dropped from 705ms to an average of 1.44ms and a peak of 1.67ms, an improvement of approximately 490 times.

Xiaomi engineer Wenchao Hao targeted the zs_free() lock contention in the zsmalloc memory allocator, which is the core allocator for zRAM and Zswap, relied upon heavily when memory is tight on Android phones and older devices.

After optimization, real-world testing showed a 1.83x speedup in concurrent memory release on a Raspberry Pi 4B and a 1.4x speedup on a 20-core Intel server.

Following the kernel community's cadence, there is typically only one LTS version per year, and 7.3 is the most favored candidate this year. If it is selected as LTS, the performance patches from ByteDance, ZTE, and Xiaomi will be maintained long-term by major distributions for several years.