~/net-history Six decades of networking, told through classic papers cd ../blog
1960s–1970s

The Birth of Packet Switching

Cold-War resilience concerns and time-sharing computers converge: data no longer needs a dedicated circuit — it can travel in packets.

4 milestones
A RAND Production
1964
On Distributed Communications Networks
Paul Baran
RANDREPORT
1964 · RAND Report

Proposed breaking messages into blocks routed independently over a redundant mesh — packet switching before the name existed, motivated by surviving nuclear attack.

OPEN MY NOTES →
Historical Event
1969
ARPANET Sends Its First Message
UCLA → SRI
EVENTARPANET
1969 · ARPANET, historical event

UCLA attempts to send "LOGIN" to SRI; the system crashes after "LO". The internet's first message is, fittingly, truncated.

OPEN MY NOTES →
An IEEE Picture
1974
A Protocol for Packet Network Intercommunication
Vinton Cerf, Robert Kahn
IEEE T-COMTCP
1974 · IEEE Transactions on Communications

The TCP paper. Introduced gateways and a common host-to-host protocol so heterogeneous networks could interconnect — the intellectual foundation of the Internet.

OPEN MY NOTES →
A Xerox PARC Production
1976
Ethernet: Distributed Packet Switching for Local Computer Networks
Robert Metcalfe, David Boggs
CACMLAN
1976 · Communications of the ACM

A shared coaxial cable, carrier sensing, and randomized backoff. Fifty years later, the name survives every technology inside it being replaced.

OPEN MY NOTES →
1980s

Principles & Protocols

The Internet gets its architecture — and its philosophy. The decade that decided what belongs in the network and what belongs at the ends.

6 milestones
Historical Event
1981
RFC 791 & 793 Standardize IP and TCP
Jon Postel, ed.
EVENTRFC
1981 · IETF RFC series

Postel's specifications split the stack into IP and TCP. On January 1, 1983 ("flag day"), ARPANET cuts over to TCP/IP for good.

OPEN MY NOTES →
An MIT Production
1984
End-to-End Arguments in System Design
Jerome Saltzer, David Reed, David Clark
ACM TOCSE2E
1984 · ACM Transactions on Computer Systems

Functions should live at the endpoints unless the network can do them completely and cheaply. The most-cited design principle in networking — and the most argued about.

OPEN MY NOTES →
A SIGCOMM Production
1988
Development of the Domain Name System
Paul Mockapetris, Kevin Dunlap
SIGCOMMDNS
1988 · ACM SIGCOMM

Replaced a single HOSTS.TXT file with a delegated, hierarchical, cached distributed database. Still resolving your queries today.

OPEN MY NOTES →
A SIGCOMM Production
1988
Congestion Avoidance and Control
Van Jacobson, Michael Karels
SIGCOMMCC
1988 · ACM SIGCOMM

Written after congestion collapse nearly froze the 1986 Internet. Slow start, AIMD, RTT estimation — the congestion control canon, and arguably the paper that saved the Internet.

OPEN MY NOTES →
A SIGCOMM Production
1988
The Design Philosophy of the DARPA Internet Protocols
David Clark
SIGCOMMARCH
1988 · ACM SIGCOMM

A retrospective on why the Internet looks the way it does: survivability first, accountability last. Explains every design wart you've ever cursed at.

OPEN MY NOTES →
A SIGCOMM Production
1989
Analysis and Simulation of a Fair Queueing Algorithm
Alan Demers, Srinivasan Keshav, Scott Shenker
SIGCOMMFQ
1989 · ACM SIGCOMM

Made per-flow fairness a router primitive. The intellectual ancestor of every scheduling and QoS mechanism since.

OPEN MY NOTES →
1990s

Scaling the Internet

The web arrives and traffic explodes. The decade of keeping a research network alive as it becomes global infrastructure.

4 milestones
Historical Event
1991
The World Wide Web Goes Public
Tim Berners-Lee, CERN
EVENTWWW
1991 · CERN, historical event

Berners-Lee releases WWW from CERN. Within five years, the web transforms the Internet from an academic tool into society's communication substrate — and its traffic pattern.

OPEN MY NOTES →
An IEEE/ACM Picture
1993
Random Early Detection Gateways for Congestion Avoidance
Sally Floyd, Van Jacobson
ToNRED
1993 · IEEE/ACM Transactions on Networking

Routers should drop packets before queues fill. RED opened the field of active queue management — a debate still running through CoDel and PIE today.

OPEN MY NOTES →
Historical Event
1994
CIDR and BGP-4 Deploy
IETF
EVENTBGP
1994 · IETF, historical event

Classless addressing and route aggregation rescue the Internet from address exhaustion and routing-table explosion. BGP-4 remains the Internet's routing protocol — for better or worse.

OPEN MY NOTES →
An SOSP Production
1999
The Click Modular Router
Eddie Kohler, Robert Morris, Benjie Chen, et al.
SOSPSW-DP
1999 · ACM SOSP

A router as a graph of composable software elements. Made packet processing programmable and inspired two decades of software dataplanes.

OPEN MY NOTES →
2000s

Overlays & Rethinking Control

If the network can't change, build on top of it — or take its control plane away. P2P systems flourish, and SDN is born.

7 milestones
A SIGCOMM Production
2001
Chord: A Scalable Peer-to-peer Lookup Service
Ion Stoica, Robert Morris, David Karger, et al.
SIGCOMMDHT
2001 · ACM SIGCOMM

Consistent hashing on a ring: O(log N) lookup with no central authority. The DHT abstraction behind everything from BitTorrent trackers to Dynamo-style storage.

OPEN MY NOTES →
A SIGCOMM Production
2002
Congestion Control for High Bandwidth-Delay Product Networks
Dina Katabi, Mark Handley, Charlie Rohrs
SIGCOMMXCP
2002 · ACM SIGCOMM

Let routers tell senders exactly how to adjust. XCP never deployed, but explicit-feedback thinking echoes through DCTCP, HPCC, and modern datacenter transports.

OPEN MY NOTES →
A CCR Feature
2005
A Clean Slate 4D Approach to Network Control and Management
Albert Greenberg, Gisli Hjalmtysson, David Maltz, et al.
CCR4D
2005 · ACM SIGCOMM CCR

Argued for ripping decision logic out of routers into a logically centralized plane. The blueprint SDN would follow.

OPEN MY NOTES →
A SIGCOMM Production
2007
Ethane: Taking Control of the Enterprise
Martín Casado, Michael Freedman, Justin Pettit, et al.
SIGCOMMPRE-SDN
2007 · ACM SIGCOMM

Centralized policy, dumb switches, flow-level control — deployed in a real campus network. The direct precursor to OpenFlow, and to Nicira.

OPEN MY NOTES →
A CCR Feature
2008
OpenFlow: Enabling Innovation in Campus Networks
Nick McKeown, Tom Anderson, Hari Balakrishnan, et al.
CCRSDN
2008 · ACM SIGCOMM CCR

A pragmatic compromise — expose the flow table, keep the hardware. Turned software-defined networking from a position paper into an industry.

OPEN MY NOTES →
A SIGCOMM Production
2008
A Scalable, Commodity Data Center Network Architecture
Mohammad Al-Fares, Alexander Loukissas, Amin Vahdat
SIGCOMMFAT-TREE
2008 · ACM SIGCOMM

Fat-trees built from cheap merchant silicon can match big-iron routers. The paper that made Clos topologies the default language of datacenter networking.

OPEN MY NOTES →
A SIGCOMM Production
2009
VL2: A Scalable and Flexible Data Center Network
Albert Greenberg, James Hamilton, Navendu Jain, et al.
SIGCOMMVL2
2009 · ACM SIGCOMM

Valiant load balancing plus flat addressing gives every service the illusion of one big switch. Microsoft's answer to datacenter agility, informed by real traffic measurement.

OPEN MY NOTES →
2010s

The Datacenter Decade

Hyperscalers become the center of gravity. Networks get custom transports, programmable chips, and are co-designed with the applications they serve.

8 milestones
A SIGCOMM Production
2010
Data Center TCP (DCTCP)
Mohammad Alizadeh, Albert Greenberg, David Maltz, et al.
SIGCOMMDCTCP
2010 · ACM SIGCOMM

Use ECN marks as a fine-grained congestion signal to keep queues near-empty. Defined the datacenter transport agenda for the decade.

OPEN MY NOTES →
A Google Production
2013
B4: Experience with a Globally-Deployed Software Defined WAN
Sushant Jain, Alok Kumar, Subhasree Mandal, et al.
SIGCOMMSD-WAN
2013 · ACM SIGCOMM

SDN leaves the lab: Google runs its inter-datacenter WAN at near-100% utilization with centralized traffic engineering. Proof that the ideas worked at planetary scale.

OPEN MY NOTES →
A CCR Feature
2014
P4: Programming Protocol-Independent Packet Processors
Pat Bosshart, Dan Daly, Glen Gibb, et al.
CCRP4
2014 · ACM SIGCOMM CCR

If OpenFlow opened the flow table, P4 opened the parser and pipeline. Switches become compile targets — enabling in-network computing from telemetry to consensus offload.

OPEN MY NOTES →
A Google Production
2015
Jupiter Rising: A Decade of Clos Topologies and Centralized Control
Arjun Singh, Joon Ong, Amit Agarwal, et al.
SIGCOMMJUPITER
2015 · ACM SIGCOMM

Five generations of datacenter fabric in one retrospective, scaling to 1.3 Pbps. The rare paper that shows how production networks actually evolve.

OPEN MY NOTES →
A Microsoft Production
2015
Congestion Control for Large-Scale RDMA Deployments
Yibo Zhu, Haggai Eran, Daniel Firestone, et al.
SIGCOMMDCQCN
2015 · ACM SIGCOMM

Made RoCEv2 deployable at cloud scale by taming PFC's pathologies. RDMA moves from HPC niche to the backbone of storage and, later, AI training.

OPEN MY NOTES →
A Google Production
2016
BBR: Congestion-Based Congestion Control
Neal Cardwell, Yuchung Cheng, C. Stephen Gunn, et al.
ACM QUEUEBBR
2016 · ACM Queue

Stop treating loss as the signal; model the pipe's bandwidth and RTT instead. Thirty years after Jacobson, congestion control gets its biggest rethink — deployed on YouTube.

OPEN MY NOTES →
A Google Production
2017
The QUIC Transport Protocol: Design and Internet-Scale Deployment
Adam Langley, Alistair Riddoch, Alyssa Wilk, et al.
SIGCOMMQUIC
2017 · ACM SIGCOMM

A new transport in userspace over UDP, deployed to billions by shipping it in Chrome. Escaped protocol ossification — and became HTTP/3.

OPEN MY NOTES →
A Microsoft Production
2018
Azure Accelerated Networking: SmartNICs in the Public Cloud
Daniel Firestone, Andrew Putnam, Sambhrama Mundkur, et al.
NSDISMARTNIC
2018 · USENIX NSDI

FPGA-based SmartNICs offload the host networking stack across the entire Azure fleet. Hardware-software co-design becomes the default playbook for cloud networking.

OPEN MY NOTES →
2020s

Optics, Offload & the AI Era

Moore's law slows while AI traffic explodes. The network reaches for optical switching and is co-designed with the training jobs running on top of it.

5 milestones
A Microsoft Production
2020
Sirius: A Flat Datacenter Network with Nanosecond Optical Switching
Hitesh Ballani, Paolo Costa, Raphael Behrendt, et al.
SIGCOMMOPTICS
2020 · ACM SIGCOMM

Nanosecond-granularity optical circuit switching across a flat datacenter. A glimpse of networks freed from electrical packet switches entirely.

OPEN MY NOTES →
A Google Production
2022
Jupiter Evolving: Optical Circuit Switches and SDN in Production
Leon Poutievski, Omid Mashayekhi, Joon Ong, et al.
SIGCOMMOCS
2022 · ACM SIGCOMM

Optical circuit switches replace the spine layer in production — no more fabric-wide forklift upgrades. OCS jumps from research prototype to the world's largest network.

OPEN MY NOTES →
An MIT Production
2023
TopoOpt: Co-optimizing Network Topology and Parallelization for ML
Weiyang Wang, Moein Khazraee, Zhizhen Zhong, et al.
NSDIML-NET
2023 · USENIX NSDI

Reconfigure the physical topology to match each ML job's communication pattern. The network stops being a fixed substrate and becomes part of the training system.

OPEN MY NOTES →
An Alibaba Production
2024
Alibaba HPN: A Data Center Network for LLM Training
Kun Qian, Yongqing Xi, Jiamin Cao, et al.
SIGCOMMLLM
2024 · ACM SIGCOMM

A dual-plane, rail-optimized fabric purpose-built for LLM training traffic: few, fat, bursty flows that defeat classic ECMP assumptions.

OPEN MY NOTES →
A Meta Production
2024
RDMA over Ethernet for Distributed AI Training at Meta Scale
Adithya Gangidi, Rui Miao, Shengbao Zheng, et al.
SIGCOMMAI-RDMA
2024 · ACM SIGCOMM

Operating RoCE fabrics for AI clusters with tens of thousands of GPUs. The 1988 congestion problem returns at 400 Gbps — and the story is far from over.

OPEN MY NOTES →