Congestion Control

Congestion Control provides performance isolation when multiple applications running on the same cluster. Additionally, it prevents congestion spreading when there is a slow receiver, reduce latency in the cluster, improves fairness, prevents parking-lot effects and packet's drop in lossy networks.

The diagram below shows an example of head of the line blocking scenario.

Developing a congestion control algorithm for datacenters present the following challenges:

• Several µ-sec of latency with hundreds of Gbps of bandwidth

  • Congestion buildup is fast, so the congestion loop should be short

• A wide variety of traffic types, topologies and applications

  • Hard to develop an algorithm that suits all

  • Congestion Control algorithms are constantly being introduced with new congestion indications

• Hardware implementation is not robust enough

• Software implementation reacts too slow

To face the challenges above, NVIDIA CC algorithm is developed on top of an infrastructure with the following characteristics: