Linux Kernel Innovations Powering the AI Agent Era – Highlights from China’s 20th CLK

Main Forum: AI Agent Era Linux Kernel Status and Development

Large‑model AI agents introduce four functional dimensions—perception, memory, planning, and execution—that stress kernel subsystems. The analysis identifies four primary bottlenecks: (1) scheduler latency when coordinating massive parallel inference tasks; (2) memory fragmentation and pressure caused by frequent model loading/unloading; (3) storage I/O contention for model checkpoints and streaming data; and (4) metric collection overhead for fine‑grained performance monitoring. The speaker proposes that the kernel must provide ultra‑low‑latency, high‑concurrency execution primitives, such as pre‑emptible real‑time scheduling classes, scalable memory cgroup policies, and fast path I/O paths that bypass traditional page‑cache layers.

Round‑Table Forum: Opportunities, Challenges, and Future Directions for Linux Kernel in the AI Era

Technical leads from major vendors concur that compute scheduling and memory management are the most critical challenges for AI workloads. They call for smarter resource orchestration that can dynamically bind GPUs, NPUs, and CPUs to AI agents, and for kernel‑level abstractions that expose heterogeneous accelerator capabilities while preserving isolation. The discussion emphasizes deep investment in core kernel technologies (e.g., cgroup v2, schedulers, and memory compaction) to sustain rapid AI‑driven product iteration.

Memory Management & Optimization Sub‑Forum

1. ZRAM Multi‑Compression Algorithm Efficiency

A new user‑experience metric called memory‑speed measures application start‑up latency under heavy background load. By configuring ZRAM to select among multiple compression algorithms (e.g., LZ4, ZSTD, and LZ4HC) based on runtime workload characteristics, the team achieved a 2 %–20 % reduction in app‑launch time on Android devices. The implementation uses a per‑page heuristic that prefers fast‑compressing algorithms for latency‑sensitive pages and higher‑ratio algorithms for less time‑critical data.

2. Heterogeneous Compression for ZRAM Using GPU Acceleration

The compression pipeline is offloaded to the GPU to eliminate CPU‑GPU contention. Data is batched in contiguous buffers and transferred via a zero‑copy DMA mapping, allowing the GPU to compress pages without intermediate copies. This frees CPU cycles for latency‑sensitive UI tasks. Benchmarks show a 30 % reduction in CPU utilization during peak memory pressure while maintaining comparable compression ratios.

3. ZCACHE Asynchronous File‑Compression Cache Management

The original Zcache suffered from slow reclaim and low compression efficiency due to synchronous compression and the ZBud allocator. The enhanced design introduces an asynchronous compression worker pool and a per‑zone cache allocator that decouples page reclamation from compression. In multi‑background Android scenarios, the optimized ZCACHE improves overall app launch speed by an average of 12.1 %.

4. Uncached Buffer I/O in F2FS

Uncached buffer I/O bypasses the page‑cache LRU management, treating I/O buffers as “fire‑and‑forget” objects. The implementation adds a new F2FS I/O path that allocates buffers directly from a slab cache and skips page‑cache insertion. Real‑world tests reduced Page Cache memory usage from 5.5 GB to 200 MB and lowered ks​wapd activity from 55 % to near zero, while still permitting concurrent reads/writes through the existing page‑cache for non‑uncached traffic.

5. EROFS Direct I/O for Compressed Files

To avoid page‑cache pressure from read‑only AI model files, a direct‑I/O path was added to EROFS. The path decompresses data directly into user‑space buffers, incorporates ztail‑packing, fragment handling, and deduplication, and does not require disk‑aligned I/O. In low‑memory conditions, this approach yields a 54.6 % read‑performance improvement over traditional buffered I/O.

File System & Storage Sub‑Forum

1. Parallelizing Filesystem Writeback

A multi‑threaded writeback scheme was implemented for F2FS. Writeback threads are pinned to separate CPU cores and coordinate via a per‑device workqueue, allowing concurrent flushing of dirty pages. On a smartphone test platform, writeback throughput increased by 22 % during short‑burst high‑load writes.

2. Host‑Initiated Defragmentation (HID)

HID enables the host to trigger on‑the‑fly defragmentation of fragmented files. The workflow includes scanning for fragmented extents, issuing asynchronous rewrite commands, and updating the file’s extent map without blocking I/O. Since its proposal in 2019, HID has been upstreamed and demonstrated >300 % speedup for copying 2 GB files in heavily fragmented UFS 4.1 environments.

3. Zoned Storage Performance Exploration

Zoned Storage (ZNS) moves zone management into the OS layer, allowing the filesystem to respect zone write‑order and reduce write amplification. The presented optimizations focus on garbage‑collection (GC) efficiency—using zone‑aware victim selection and background GC threads—and on copy‑offloading techniques that move data between zones without CPU involvement. Early results show latency‑controlled random‑read performance and a measurable reduction in write amplification.

Scheduling, Performance & Debugging Sub‑Forum

Energy‑Efficient I/O: Block‑Level Device Frequency Constraint

A PM‑QoS‑based mechanism constrains block‑device frequency scaling. The kernel monitors I/O queue depth and, when the depth falls below a threshold, raises the device’s operating frequency to meet latency‑sensitive request deadlines. Experiments indicate a ~15 % reduction in command‑completion latency under intermittent I/O and a ~5 % bandwidth increase under sustained load.

AI Infrastructure & eBPF Application Sub‑Forum

cpuidle_ext Framework Based on eBPF for Custom Low‑Power Policies

The cpuidle_ext framework extends the standard cpuidle governor with BPF struct_ops. User‑space programs can load BPF programs that inspect runtime metrics (e.g., audio playback state) and dynamically select idle states. A demonstration reduced CPU C0 occupancy from 50 % to 10 % during music playback, cutting overall system power consumption by 5 %.