13 papers selected.
Empowering WebAssembly with Thin Kernel Interfaces
Arjun Ramesh, Tianshu Huang, Ben L. Titzer, Anthony Rowe 0001
TL;DR — A new OS interface design exposes thin, capability-based kernel primitives directly to WebAssembly modules, eliminating the POSIX translation layer.
Why notable — WebAssembly is increasingly used beyond the browser as a portable, sandboxed compute substrate; this work shows that rethinking the system interface from scratch yields significantly lower overhead and better safety properties than layering Wasm on top of POSIX.
Revealing the Unstable Foundations of eBPF-Based Kernel Extensions
Shawn Wanxiang Zhong, Jing Liu 0074, Andrea C. Arpaci-Dusseau, Remzi H. Arpaci-Dusseau
TL;DR — A systematic study exposes how eBPF programs silently break across kernel versions due to undocumented and unstable kernel data structure dependencies.
Why notable — From the Arpaci-Dusseau group, this paper delivers surprising and practically consequential findings: a large fraction of real-world eBPF programs are fragile across kernel versions, undermining the widely held assumption that eBPF is a safe and stable extension mechanism.
Pegasus: Transparent and Unified Kernel-Bypass Networking for Fast Local and Remote Communication
Dinglan Peng, Congyu Liu, Tapti Palit, Anjo Vahldiek-Oberwagner et al.
TL;DR — Pegasus provides a single kernel-bypass networking stack that transparently accelerates both intra-host (IPC) and inter-host communication without application changes.
Why notable — Unifying local and remote fast paths is a long-standing challenge; Pegasus demonstrates that the same RDMA-style techniques can be applied to loopback traffic, yielding substantial latency reductions for microservice workloads with no API changes.
Daredevil: Rescue Your Flash Storage from Inflexible Kernel Storage Stack
Junzhe Li, Ran Shu 0001, Jiayi Lin 0007, Qingyu Zhang 0005 et al.
TL;DR — Daredevil bypasses the rigid Linux block layer to allow flash storage devices to express fine-grained I/O semantics directly to applications.
Why notable — The Linux storage stack was designed for spinning disks and consistently imposes unnecessary overhead on modern NVMe SSDs; Daredevil demonstrates that rethinking the kernel/device interface boundary unlocks significant throughput and latency gains that the existing stack structurally prevents.
Towards Efficient Flash Caches with Emerging NVMe Flexible Data Placement SSDs
Michael Allison, Arun George, Javier González 0006, Dan Helmick et al.
TL;DR — This paper shows how NVMe Flexible Data Placement (FDP) SSDs can be exploited by flash cache software to dramatically reduce write amplification and improve device lifetime.
Why notable — FDP is a newly standardized NVMe feature; this is one of the first systems papers to demonstrate end-to-end integration with a production-grade flash caching stack, revealing concrete performance and endurance benefits that motivate wider adoption.
Chrono: Meticulous Hotness Measurement and Flexible Page Migration for Memory Tiering
Zhenlin Qi, Shengan Zheng, Ying Huang, Yifeng Hui et al.
TL;DR — Chrono introduces fine-grained, low-overhead hotness tracking and a flexible page migration policy that adapts to workload dynamics for tiered memory systems.
Why notable — As CXL-based memory tiering becomes a reality in data centers, accurate hotness estimation is critical; Chrono’s approach substantially outperforms existing kernel mechanisms and sets a new baseline for OS-level tiered memory management.
Adios to Busy-Waiting for Microsecond-scale Memory Disaggregation
Wonsup Yoon, Jisu Ok, Sue Moon, Youngjin Kwon
TL;DR — This work eliminates CPU-wasting busy-waiting in disaggregated memory systems by designing interrupt-driven mechanisms that still meet microsecond latency targets.
Why notable — Busy-waiting is the conventional wisdom for achieving low latency in disaggregated memory, yet it burns entire CPU cores; this paper challenges that assumption and shows interrupt-based designs can match latency while freeing substantial compute, which matters greatly at scale.
Collaborative Text Editing with Eg-walker: Better, Faster, Smaller
Joseph Gentle, Martin Kleppmann
TL;DR — Eg-walker is a new CRDT algorithm for collaborative text editing that is simultaneously faster, more memory-efficient, and produces smaller operation logs than prior state-of-the-art CRDTs.
Why notable — Collaborative editing CRDTs have been considered a mature area, making it surprising that Eg-walker achieves order-of-magnitude improvements across all key metrics; the result, from Martin Kleppmann, will likely become the new reference design for replicated text data structures.
Ladon: High-Performance Multi-BFT Consensus via Dynamic Global Ordering
Hanzheng Lyu, Shaokang Xie, Jianyu Niu, Chen Feng 0001 et al.
TL;DR — Ladon achieves high throughput in Byzantine fault-tolerant consensus by running multiple BFT instances in parallel and dynamically merging their outputs into a consistent global order.
Why notable — Byzantine consensus is notoriously throughput-limited; Ladon’s multi-instance approach with a novel global ordering layer demonstrates near-linear throughput scaling with the number of consensus instances, a significant advance for permissioned blockchain and critical-infrastructure scenarios.
HybridFlow: A Flexible and Efficient RLHF Framework
Guangming Sheng, Chi Zhang 0022, Zilingfeng Ye, Xibin Wu et al.
TL;DR — HybridFlow introduces a hybrid dataflow model for RLHF training that co-schedules the actor, critic, and reward models to maximize GPU utilization across heterogeneous cluster configurations.
Why notable — RLHF has become central to LLM alignment yet existing frameworks map poorly to its multi-model, tightly coupled training loop; HybridFlow’s design substantially improves end-to-end training throughput and provides a principled abstraction for future alignment training research.
CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion
Jiayi Yao, Hanchen Li, Yuhan Liu 0004, Siddhant Ray et al.
TL;DR — CacheBlend reuses KV caches from multiple pre-computed document chunks and fuses them selectively at inference time, avoiding the quadratic cost of full re-encoding for retrieval-augmented generation.
Why notable — RAG is a dominant LLM deployment pattern, but cache reuse across dynamically assembled contexts is unsolved; CacheBlend’s selective fusion strategy delivers large time-to-first-token reductions with negligible quality degradation, directly improving the economics of production LLM serving.
AlloyStack: A Library Operating System for Serverless Workflow Applications
Jianing You, Kang Chen, Laiping Zhao, Yiming Li et al.
TL;DR — AlloyStack is a library OS tailored for serverless workflows that collapses function boundaries within a workflow into a single address space to eliminate inter-function communication overhead.
Why notable — Serverless workflows suffer from high invocation and communication latency because each function is an isolated container; AlloyStack’s library OS approach is a principled architectural answer that shows substantial end-to-end latency and cost improvements for real workflow benchmarks.
CRAVE: Analyzing Cross-Resource Interaction to Improve Energy Efficiency in Systems-on-Chip
Dipayan Mukherjee, Sam Hachem, Jeremy Bao, Curtis Madsen et al.
TL;DR — CRAVE models the cross-resource interference between CPU, GPU, and memory subsystems on SoCs to guide software-level energy optimization decisions.
Why notable — Energy efficiency is increasingly a first-class constraint in both mobile and data center SoCs, yet interactions between on-chip resources are poorly understood at the software level; CRAVE’s analysis framework reveals counter-intuitive interference patterns and enables measurable energy savings without hardware changes.