Hello World.

Papers of Interest in SIGCOMM 22

SIGCOMM 22 contains 11 technical sessions and 55 papers are accepted.

  1. Datacenter Networking
  2. 5G Networks
  3. Congestion Control
  4. Wide Area Networks
  5. Testing and Verification (most of them mentioned programmable dataplane)
  6. Machine Learning
  7. Monitoring and Measurement
  8. Sensing and Wireless Communication
  9. Programmable Data Planes
    • Predictable vFabric on Informative Data Plane
    • Using trio: juniper networks’ programmable chipset - for emerging in-network applications
    • Thanos: Programmable Multi-Dimensional Table Filters for Line Rate Network Functions (hard core)
    • Stateful Multi-Pipelined Programmable Switches (same author as the above paper)
    • FAst In-Network GraY Failure Detection for ISPs
  10. Denial of Service Defense and Storage Networks
  11. Host Networking and Video Delivery

Interested Papers

Jupiter Evolving: Transforming Google’s Datacenter Network via Optical Circuit Switches and Software-Defined Networking

Google proposed a new datacenter architecture in this paper.

  • A datacenter interconnection layer
  • Centralized SDN controller for traffic engineering
  • Automated network operations for topology engineering

The combination of traffic and topology engineering on direct-connect fabrics achieves simliar throughput as Clos fabrics for Google’s production traffic patterns.

The paper targets the challenge of incremental refresh of the network infrastructure. Since the typical Clos network requires pre-building spine at the maximum-scale, which may derated the bandiwidth of links to aggregation block.

  1. spine层一旦建好,难以修改,因为这需要很大的成本
  2. spine层的带宽会限制后续更新的其他模块带宽,造成计算和网络资源失配


Aequitas: Admission Control for Performance-Critical RPCs in Datacenters

Another paper of Google, targeting guarteeing SLO of RPCs.

Traditional traffic classification categories cannot sufficiently capture their importance due to wide variations in RPC characteristics.

As of 2021, RPCs generate 95%+ of the application traffic in Google production datacenters, of which ∼75% is to and from storage systems.

An RPC is a programmatic request for action or information between components of applications, and it can consist of dozens of individual packets. Hundreds of RPCs can be on the critical path to completing an application-level operation.

Cebinae: Scalable In-network Fairness Augmentation

Enabling the fairness among different congestion control protocols with programmable switches.


Key insights

  1. High-level flow scheduling is enough: > it suffices for the network to redistribute bandwidth from flows that have met/exceeded their fair share to flows that have not.
  2. Advanced scheduling logic can be adopted based on the above simplification.

Open sourced in https://github.com/eniac/Cebinae

FlyMon: Enabling On-the-Fly Task Reconfiguration for Network Measurement


Sketch-based measurement system that can make on-the-fly reconfigurations on a large set of measurement tasks.

This paper may be useful for my future research.

, — Sep 3, 2022