☀️

Hello World.

In-network Computing Research

In-network computing adopts programmable network devices (e.g., programmable switches) to perform complex operations, while forwarding packets.

This post presents some representative works in the field of in-network computing.

In-network Aggregation

See this post for details.

In-network Cache

Representative group: Xin Jin.

  • SOSP 17, NetCache: Balancing Key-Value Stores with Fast In-Network Caching
  • SIGCOMM 20, NetLock: Fast, Centralized Lock Management Using Programmable Switches
  • NSDI 18, NetChain: Scale-Free Sub-RTT Coordination

In-network Virtual Functions

This area of works adopt programmable switches (or smart NICs) to implement the hardware gateway, to improve the traffic throughput in the cloud.

  • SIGCOMM 17, SilkRoad: Making Stateful Layer-4 Load Balancing Fast and Cheap Using Switching ASICs
  • HotNets 19, Accelerated Service Chaining on a Single Switch ASIC
  • SIGCOMM 21, Sailfish: Accelerating Cloud-Scale Multi-Tenant Multi-Service Gateways with Programmable Switches
  • NSDI 22, Elixir: A High-performance and Low-cost Approach to Managing Hardware/Software Hybrid Flow Tables Considering Flow Burstiness
  • NSDI 22, Tiara: A Scalable and Efficient Hardware Acceleration Architecture for Stateful Layer-4 Load Balancing

In-network Measurement

In-network measurement is usually achieved by implementing sketch algorithms with programmable switches, to perform fine-grained measurement.

  • NSDI 21, Toward Nearly-Zero-Error Sketching via Compressive Sensing
  • SIGCOMM 19, BeauCoup: Answering Many Network Traffic Queries, One Memory Update at a Time
  • SIGCOMM 20, Flow Event Telemetry on Programmable Data Plane
  • SIGCOMM 22, FlyMon: Enabling On-the-Fly Task Reconfiguration for Network Measurement
  • NSDI 23, Sketchovsky: Enabling Ensembles of Sketches on Programmable Switches

Others

  • SIGCOMM 23, NetClone: Fast, Scalable, and Dynamic Request Cloning for Microsecond-Scale RPCs
  • NSDI 24, THC: Accelerating Distributed Deep Learning Using Tensor Homomorphic Compression

— Dec 22, 2023