--- title: NSDI 2024 Digest venue: NSDI year: 2024 date: '2024-01-01' tags: [] paper_count: 13 draft: false --- 13 papers selected. --- ### MegaScale: Scaling Large Language Model Training to More Than 10, 000 GPUs *Ziheng Jiang, Haibin Lin, Yinmin Zhong, Qi Huang *et al.** **TL;DR** — ByteDance's full production account of training LLMs at 10,000+ GPUs, with novel co-design of the network stack, fault tolerance, and collective communication to sustain near-linear scaling. --- ### Harmony: A Congestion-free Datacenter Architecture *Saksham Agarwal, Qizhe Cai, Rachit Agarwal 0001, David B. Shmoys *et al.** **TL;DR** — Proposes rethinking datacenter fabrics to eliminate congestion by construction rather than managing it reactively, achieving line-rate throughput without per-packet feedback from Cornell and Google. --- ### DINT: Fast In-Kernel Distributed Transactions with eBPF *Yang Zhou 0008, Xingyu Xiang, Matthew Kiley, Sowmya Dharanipragada *et al.** **TL;DR** — Demonstrates that eBPF programs executing entirely inside the kernel can enforce linearizable, ACID-compliant distributed transactions at dramatically lower latency than user-space approaches. --- ### Making Kernel Bypass Practical for the Cloud with Junction *Joshua Fried, Gohar Irfan Chaudhry, Enrique Saurez, Esha Choukse *et al.** **TL;DR** — Junction (MIT + Microsoft) is the first system to bring full kernel-bypass networking to multi-tenant cloud VMs without requiring application modifications or sacrificing isolation. --- ### SIEVE is Simpler than LRU: an Efficient Turn-Key Eviction Algorithm for Web Caches *Yazhuo Zhang, Juncheng Yang, Yao Yue, Ymir Vigfusson *et al.** **TL;DR** — Surprising finding that a single-queue eviction policy requiring almost no state changes outperforms LRU and all modern approximations on production web-cache workloads. --- ### A large-scale deployment of DCTCP *Abhishek Dhamija, Balasubramanian Madhavan, Hechao Li, Jie Meng *et al.** **TL;DR** — A rare, candid production report from Meta on deploying DCTCP at hyperscale, surfacing unexpected interactions with heterogeneous hardware, cross-traffic, and operational constraints. --- ### Brain-on-Switch: Towards Advanced Intelligent Network Data Plane via NN-Driven Traffic Analysis at Line-Speed *Jinzhu Yan, Haotian Xu, Zhuotao Liu, Qi Li 0002 *et al.** **TL;DR** — Executes neural-network inference for traffic classification directly in the programmable switch data plane at line rate, eliminating the round-trip to a CPU-based classifier. --- ### Horus: Granular In-Network Task Scheduler for Cloud Datacenters *Parham Yassini, Khaled Diab 0001, Saeed Mahloujifar, Mohamed Hefeeda* **TL;DR** — Offloads microsecond-granularity task scheduling decisions into programmable switches, reducing load-balancing latency by orders of magnitude compared to software schedulers. --- ### CAPA: An Architecture For Operating Cluster Networks With High Availability *Bingzhe Liu, Colin Scott, Mukarram Tariq, Andrew D. Ferguson *et al.** **TL;DR** — Google's production architecture for maintaining five-nines availability in Jupiter-scale cluster networks, detailing how control-plane redundancy and fast failover are achieved in practice. --- ### Revisiting Congestion Control for Lossless Ethernet *Yiran Zhang, Qingkai Meng 0001, Chaolei Hu, Fengyuan Ren* **TL;DR** — Identifies fundamental flaws in PFC-based lossless Ethernet that cause cascading head-of-line blocking and proposes a redesigned congestion control that avoids them. --- ### Sirius: Composing Network Function Chains into P4-Capable Edge Gateways *Jiaqi Gao, Jiamin Cao, Yifan Li, Mengqi Liu 0001 *et al.** **TL;DR** — Alibaba's production system for composing arbitrary NF chains into P4-programmed edge gateways, replacing a fleet of dedicated middleboxes and cutting per-packet processing cost significantly. --- ### CASSINI: Network-Aware Job Scheduling in Machine Learning Clusters *Sudarsanan Rajasekaran, Manya Ghobadi, Aditya Akella* **TL;DR** — Shows that ignoring network topology when scheduling ML training jobs causes severe collective-communication contention, and that topology-aware co-scheduling yields substantial throughput gains. --- ### Cloudcast: High-Throughput, Cost-Aware Overlay Multicast in the Cloud *Sarah Wooders, Shu Liu, Paras Jain 0001, Xiangxi Mo *et al.** **TL;DR** — Berkeley/Penn system that builds overlay multicast trees across cloud regions optimized for egress cost, enabling high-throughput data dissemination at a fraction of unicast cloud egress fees.