13 papers selected.
ASTERINAS: A Linux ABI-Compatible, Rust-Based Framekernel OS with a Small and Sound TCB
Yuke Peng, Hongliang Tian, Junyang Zhang, Ruihan Li et al.
TL;DR — A production-grade OS kernel written in Rust that exposes a full Linux ABI while confining unsafe code to a small, formally-audited framekernel core.
Why notable — ASTERINAS demonstrates that Linux compatibility and memory-safety guarantees are not mutually exclusive — unsafe Rust is isolated to under 5 kloc of framework code, giving systems operators a credible path toward a safer Linux-compatible kernel without sacrificing application portability.
Rex: Closing the language-verifier gap with safe and usable kernel extensions
Jinghao Jia, Ruowen Qin, Milo Craun, Egor Lukiyanov et al.
TL;DR — Rex introduces a new kernel-extension framework that replaces eBPF’s in-kernel verifier with a Rust-typed, LLVM-based toolchain to safely express programs that eBPF currently rejects.
Why notable — The eBPF verifier’s conservatism silently limits what practitioners can implement; Rex shows how a language-level safety guarantee can replace ad-hoc bytecode verification without changing the kernel ABI, opening the door to far richer kernel extensions.
PageFlex: Flexible and Efficient User-space Delegation of Linux Paging Policies with eBPF
Anil Yelam, Kan Wu, Zhiyuan Guo, Suli Yang et al.
TL;DR — PageFlex lets applications plug in custom page-replacement and allocation policies via eBPF hooks without modifying the kernel, achieving performance competitive with kernel-native policies.
Why notable — Memory management policy has historically been locked inside the kernel; PageFlex’s eBPF delegation mechanism gives cloud operators a principled way to tailor paging behavior per workload, directly addressing the one-size-fits-all limitation of the Linux page allocator.
μEFI: A Microkernel-Style UEFI with Isolation and Transparency
Le Chen, Yiyang Wu, Jinyu Gu 0001, Yubin Xia et al.
TL;DR — μEFI restructures UEFI firmware around microkernel principles so that individual UEFI drivers are isolated from each other and from the boot-time trusted computing base.
Why notable — Firmware vulnerabilities are notoriously hard to patch and can persist through OS reinstalls; μEFI’s approach substantially reduces the blast radius of a compromised UEFI driver while remaining compatible with existing UEFI software, making it directly relevant to secure-boot infrastructure.
Z-LFS: A Zoned Namespace-tailored Log-structured File System for Commodity Small-zone ZNS SSDs
Inhwi Hwang, Sangjin Lee 0003, Sunggon Kim, Hyeonsang Eom et al.
TL;DR — Z-LFS is an LFS designed around the tight zone-size constraints of commodity ZNS SSDs, using fine-grained segment management and zone-aware garbage collection to avoid the capacity and write-amplification pitfalls of existing approaches.
Why notable — ZNS SSDs offer significant cost and endurance advantages but mainstream file systems waste capacity on small-zone devices; Z-LFS shows that rethinking LFS segment layout specifically for small zones yields competitive throughput with substantially lower write amplification.
Crash Consistency in Block-Level Caching Systems: An Open CAS Case Study
Shaohua Duan, Youmin Chen
TL;DR — A systematic study of crash-consistency bugs in the widely deployed Open CAS block-layer cache, revealing a class of ordering violations that can silently corrupt data on unexpected power loss.
Why notable — Block-level caches are invisible to file systems and often assumed to be transparent, making these bugs particularly insidious; the paper’s taxonomy and detection methodology are directly actionable for operators running NVMe caching in production storage stacks.
HotRAP: Hot Record Retention and Promotion for LSM-trees with Tiered Storage
Jiansheng Qiu, Fangzhou Yuan, Mingyu Gao 0001, Huanchen Zhang
TL;DR — HotRAP adds a retention-and-promotion layer to LSM-tree compaction that keeps frequently accessed records in faster storage tiers by tracking access heat across compaction boundaries.
Why notable — Tiered storage deployments with LSM engines (RocksDB, LevelDB) routinely see hot data demoted to slow tiers during compaction; HotRAP’s lightweight heat tracking improves read latency by up to 5x on skewed workloads without changing the external LSM API.
TGW: Operating an Efficient and Resilient Cloud Gateway at Scale
Yifan Yang 0009, Lin He 0004, Jiasheng Zhou, Xiaoyi Shi et al.
TL;DR — TGW describes the architecture, traffic engineering, and operational lessons of a production cloud gateway handling hundreds of Tbps at a major cloud provider.
Why notable — Production-scale gateway papers with real traffic data are rare; TGW’s account of how to sustain sub-millisecond failover and linear-scale throughput under adversarial traffic patterns provides a concrete reference design for anyone building or operating large-scale edge infrastructure.
SwCC: Software-Programmable and Per-Packet Congestion Control in RDMA Engine
Hongjing Huang, Jie Zhang 0081, Xuzheng Chen, Ziyu Song et al.
TL;DR — SwCC embeds a programmable per-packet congestion-control engine directly inside an RDMA NIC, allowing operators to deploy and hot-swap CC algorithms without CPU involvement or ASIC redesign.
Why notable — RDMA congestion control has historically been frozen in NIC firmware, forcing cluster-wide firmware upgrades to try new algorithms; SwCC’s programmable datapath brings the velocity of software-defined networking to the RDMA layer at near-line-rate performance.
Opening Up Kernel-Bypass TCP Stacks
Shinichi Awamoto, Michio Honda
TL;DR — A framework that exposes kernel-bypass TCP (DPDK-based) stacks to unmodified POSIX applications by transparently interposing at the syscall level, without requiring application changes or root privileges.
Why notable — Kernel-bypass networking has been limited to purpose-built applications; this work’s zero-modification deployment model makes microsecond-latency TCP accessible to the broad ecosystem of existing networked software, which has significant practical implications for latency-sensitive cloud services.
Accelerating Nested Virtualization with HyperTurtle
Ori Ben Zur, Jakob Krebs, Shai Aviram Bergman, Mark Silberstein
TL;DR — HyperTurtle reduces the performance penalty of nested virtualization by selectively forwarding L2 guest hypercalls directly to the L0 hypervisor, bypassing the L1 intermediary for common fast paths.
Why notable — Nested virtualization is increasingly important for confidential computing and cloud-in-cloud deployments, but the overhead is often prohibitive; HyperTurtle’s selective bypass approach cuts nested VM overhead by up to 60% on I/O-intensive workloads with no guest modifications.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider
Jiahao Wang, Jinbo Han, Xingda Wei, Sijie Shen et al.
TL;DR — A production measurement study of LLM KV-cache behavior across a large cloud fleet, revealing access patterns and reuse characteristics that inform a redesigned caching policy reducing GPU memory pressure significantly.
Why notable — As LLM serving becomes a dominant cloud workload, KV-cache management is a critical bottleneck; this paper provides the first at-scale empirical characterization of KV-cache reuse in a real deployment, and its findings directly shaped policy changes that improved cache hit rates by over 30%.
GREYHOUND: Hunting Fail-Slows in Hybrid-Parallel Training at Scale
Tianyuan Wu, Wei Wang 0030, Yinghao Yu, Siran Yang et al.
TL;DR — GREYHOUND is a runtime monitoring system that automatically detects and isolates fail-slow stragglers in large-scale hybrid-parallel LLM training jobs before they degrade the entire training run.
Why notable — Fail-slow faults are notoriously harder to detect than fail-stop failures and can silently extend training jobs by hours; GREYHOUND’s production deployment demonstrates that fine-grained per-layer timing signals can catch slow nodes within seconds, making it a practical reliability tool for anyone running large training clusters.