178 lines
10 KiB
Markdown
178 lines
10 KiB
Markdown
---
|
|
title: NSDI 2025 Digest
|
|
venue: NSDI
|
|
year: 2025
|
|
date: '2025-04-28'
|
|
tags:
|
|
- networking
|
|
- distributed-systems
|
|
- cloud
|
|
- programmable-data-planes
|
|
- network-verification
|
|
- datacenter
|
|
- transport
|
|
- ml-systems
|
|
- 5g
|
|
paper_count: 13
|
|
draft: false
|
|
---
|
|
|
|
13 papers selected.
|
|
|
|
---
|
|
|
|
### PRED: Performance-oriented Random Early Detection for Consistently Stable Performance in Datacenters
|
|
|
|
*Xinle Du, Tong Li 0014, Guangmeng Zhou, Zhuotao Liu *et al.**
|
|
|
|
**TL;DR** — PRED redesigns AQM by making drop probability a direct function of per-flow performance targets rather than queue length, eliminating the instability of classic RED in modern datacenter workloads.
|
|
|
|
**Why notable** — RED has been a cornerstone of congestion control for decades; PRED's performance-centric reformulation challenges a long-held design axiom and demonstrates significantly lower tail latency at scale. It opens the door to intent-driven AQM as a first-class primitive in datacenter switches.
|
|
|
|
[→ Read paper](https://www.usenix.org/conference/nsdi25/presentation/du)
|
|
|
|
---
|
|
|
|
### Rajomon: Decentralized and Coordinated Overload Control for Latency-Sensitive Microservices
|
|
|
|
*Jiali Xing, Akis Giannoukos, Paul Loh, Shuyue Wang *et al.**
|
|
|
|
**TL;DR** — Rajomon introduces a token-based, decentralized overload control mechanism that coordinates admission across microservice call graphs without a central bottleneck.
|
|
|
|
**Why notable** — Microservice overload propagation is a persistent pain point in production clouds; Rajomon's approach of spreading load-shedding decisions across the call graph while retaining global coherence is a practical and principled contribution. The system is evaluated on realistic cloud benchmarks and shows clear SLO improvement over prior centralized and uncoordinated schemes.
|
|
|
|
[→ Read paper](https://www.usenix.org/conference/nsdi25/presentation/xing)
|
|
|
|
---
|
|
|
|
### Unlocking ECMP Programmability for Precise Traffic Control
|
|
|
|
*Yadong Liu, Yunming Xiao, Xuan Zhang, Weizhen Dang *et al.**
|
|
|
|
**TL;DR** — This work exposes fine-grained ECMP programming interfaces that allow operators to steer individual flows through specific paths in multipath datacenter fabrics with near-zero overhead.
|
|
|
|
**Why notable** — ECMP's coarse hashing has been a limiting factor in datacenter traffic engineering for years; this paper shows that commodity switch primitives can be composed to achieve precise per-flow placement, delivering measurable improvements in load balance without topology changes.
|
|
|
|
[→ Read paper](https://www.usenix.org/conference/nsdi25/presentation/liu-yadong)
|
|
|
|
---
|
|
|
|
### eTran: Extensible Kernel Transport with eBPF
|
|
|
|
*Zhongjie Chen, Qingkai Meng 0001, ChonLam Lao, Yifan Liu *et al.**
|
|
|
|
**TL;DR** — eTran uses eBPF to let applications plug in custom transport logic—including RDMA-like zero-copy paths—directly into the Linux kernel data path without modifying kernel source.
|
|
|
|
**Why notable** — The ability to safely extend kernel networking with eBPF has been widely anticipated; eTran demonstrates that full transport protocols (not just classifiers) can be realised this way, achieving performance competitive with kernel-bypass while retaining OS isolation and deployability.
|
|
|
|
[→ Read paper](https://www.usenix.org/conference/nsdi25/presentation/chen-zhongjie)
|
|
|
|
---
|
|
|
|
### White-Boxing RDMA with Packet-Granular Software Control
|
|
|
|
*Chenxingyu Zhao, Jaehong Min, Ming Liu 0027, Arvind Krishnamurthy*
|
|
|
|
**TL;DR** — This paper decomposes RDMA semantics and re-implements them in software at packet granularity, exposing hooks that allow transport policy (retransmission, congestion, multipath) to be modified without hardware changes.
|
|
|
|
**Why notable** — RDMA's black-box nature has long frustrated datacenter operators trying to deploy custom congestion control or network telemetry; white-boxing RDMA at packet granularity is a principled answer that simultaneously improves debuggability and extensibility, with demonstrated line-rate performance.
|
|
|
|
[→ Read paper](https://www.usenix.org/conference/nsdi25/presentation/zhao-chenxingyu)
|
|
|
|
---
|
|
|
|
### NDD: A Decision Diagram for Network Verification
|
|
|
|
*Zechun Li, Peng Zhang 0011, Yichi Zhang, Hongkun Yang*
|
|
|
|
**TL;DR** — NDD is a new symbolic data structure that compactly represents the forwarding behavior of large networks, enabling network verification queries orders of magnitude faster than BDD-based predecessors.
|
|
|
|
**Why notable** — Network verification tools like Batfish and ARC rely on set representations whose size can explode on real networks; NDD's topology-aware decomposition dramatically reduces verification time for common queries (reachability, loop detection) and is likely to be adopted as a backend in production verification pipelines.
|
|
|
|
[→ Read paper](https://www.usenix.org/conference/nsdi25/presentation/li-zechun)
|
|
|
|
---
|
|
|
|
### VEP: A Two-stage Verification Toolchain for Full eBPF Programmability
|
|
|
|
*Xiwei Wu, Yueyang Feng, Tianyi Huang, Xiaoyang Lu *et al.**
|
|
|
|
**TL;DR** — VEP combines abstract interpretation with deductive verification to prove safety properties of eBPF programs that the Linux kernel verifier rejects, enabling a much wider class of programs to run safely in the kernel.
|
|
|
|
**Why notable** — The Linux eBPF verifier's conservatism blocks many useful programs; VEP's two-stage approach (lightweight in-kernel check plus offline proof) expands the programmability frontier without weakening safety guarantees, directly affecting every operator who writes eBPF for networking or observability.
|
|
|
|
[→ Read paper](https://www.usenix.org/conference/nsdi25/presentation/wu-xiwei)
|
|
|
|
---
|
|
|
|
### MTP: Transport for In-Network Computing
|
|
|
|
*Tao Ji, Rohan Vardekar, Balajee Vamanan, Brent E. Stephens *et al.**
|
|
|
|
**TL;DR** — MTP is a transport protocol co-designed with programmable switches that allows in-network compute operations (aggregation, filtering) to be expressed as first-class transport primitives with reliability and flow-control guarantees.
|
|
|
|
**Why notable** — Existing transports treat the network as a dumb pipe and retrofit in-network compute as an afterthought; MTP shows that rethinking transport abstractions around programmable hardware can yield substantial throughput gains for distributed ML and key-value workloads.
|
|
|
|
[→ Read paper](https://www.usenix.org/conference/nsdi25/presentation/ji)
|
|
|
|
---
|
|
|
|
### State-Compute Replication: Parallelizing High-Speed Stateful Packet Processing
|
|
|
|
*Qiongwen Xu, Sebastiano Miano, Xiangyu Gao, Tao Wang 0088 *et al.**
|
|
|
|
**TL;DR** — This paper introduces a replication model for stateful packet processing that partitions state across multiple pipeline replicas and reconciles updates at line rate, breaking the single-pipeline bottleneck of P4 programs.
|
|
|
|
**Why notable** — Stateful P4 programs that require per-flow counters or heavy-hitter detection are bottlenecked by single-instance state; the state-compute replication model is a general technique applicable across telemetry, load balancing, and security use cases, with a prototype running at 100 Gbps.
|
|
|
|
[→ Read paper](https://www.usenix.org/conference/nsdi25/presentation/xu-qiongwen)
|
|
|
|
---
|
|
|
|
### Efficient Direct-Connect Topologies for Collective Communications
|
|
|
|
*Liangyu Zhao, Siddharth Pal, Tapan Chugh, Weiyang Wang *et al.**
|
|
|
|
**TL;DR** — This paper derives a family of direct-connect network topologies mathematically optimized for the all-reduce and all-to-all collectives used in large-scale ML training, achieving better bandwidth and lower diameter than fat-tree alternatives at the same port count.
|
|
|
|
**Why notable** — As AI training clusters scale to thousands of GPUs, interconnect topology becomes a first-order performance determinant; the paper's analytical framework for topology synthesis fills a long-standing gap between graph theory and practical cluster networking, and the proposed topologies outperform fat-trees on real collective benchmarks.
|
|
|
|
[→ Read paper](https://www.usenix.org/conference/nsdi25/presentation/zhao-liangyu)
|
|
|
|
---
|
|
|
|
### SimAI: Unifying Architecture Design and Performance Tuning for Large-Scale Large Language Model Training with Scalability and Precision
|
|
|
|
*Xizheng Wang, Qingxu Li, Yichi Xu, Gang Lu *et al.**
|
|
|
|
**TL;DR** — SimAI is a high-fidelity simulator that jointly models compute, network, and storage interactions in LLM training clusters, enabling architects to evaluate topology and parallelism strategy choices without running full cluster experiments.
|
|
|
|
**Why notable** — Designing training infrastructure for frontier LLMs is prohibitively expensive to explore empirically; SimAI's validated accuracy against production Alibaba clusters makes it a practical tool for the community and provides rare quantitative insight into how network architecture choices affect end-to-end training throughput.
|
|
|
|
[→ Read paper](https://www.usenix.org/conference/nsdi25/presentation/wang-xizheng-simai)
|
|
|
|
---
|
|
|
|
### Minder: Faulty Machine Detection for Large-scale Distributed Model Training
|
|
|
|
*Yangtao Deng, Xiang Shi, Zhuo Jiang, Xingjian Zhang 0009 *et al.**
|
|
|
|
**TL;DR** — Minder continuously monitors collective-communication timing patterns during LLM training to pinpoint faulty nodes—including subtle slow or flapping hardware—without requiring dedicated diagnostic jobs.
|
|
|
|
**Why notable** — GPU cluster faults that cause training slowdowns rather than outright crashes are notoriously hard to attribute; Minder's passive monitoring approach catches a wide class of hardware anomalies (including network stragglers) with sub-minute detection latency, and its deployment at scale in Alibaba's training infrastructure gives the results strong operational credibility.
|
|
|
|
[→ Read paper](https://www.usenix.org/conference/nsdi25/presentation/deng)
|
|
|
|
---
|
|
|
|
### Efficient Multi-WAN Transport for 5G with OTTER
|
|
|
|
*Mary Hogan, Gerry Wan, Yiming Qiu, Sharad Agarwal *et al.**
|
|
|
|
**TL;DR** — OTTER is a multi-path transport system for 5G user-plane traffic that dynamically bonds heterogeneous WAN links (e.g., terrestrial + satellite) while hiding link asymmetry and jitter from applications.
|
|
|
|
**Why notable** — 5G deployments increasingly rely on diverse backhaul paths with wildly different characteristics; OTTER's scheduler, validated on a Microsoft production 5G deployment, demonstrates that significant throughput and reliability gains are achievable through principled multi-WAN design, informing both operator practice and future transport standards.
|
|
|
|
[→ Read paper](https://www.usenix.org/conference/nsdi25/presentation/hogan)
|
|
|