OnRamp NSDI 2021

Breaking the Transience Equilibrium Nexus A New Approach to Datacenter Packet Transport

Posted by Yiran on April 14, 2021

云数据中心拥塞控制算法: factoring of datacenter congestion control into two separate control loops, one each for transience and equilibrium.

Core idea

解耦 Transience-Equilibrium (笔者理解为瞬态拥塞和持久/稳态拥塞):

  • 传统 Congestion Cntrol 处理 Equilibrium congestion
  • 增加基于端到端延时的pause机制处理 Transience congestion

Motivation

  • Transience-Equilibrium Tension. 这是这篇论文的核心. 作者给出了一般CC的抽象, 来说明目前CC在同时处理Transience态和Equilibrium态拥塞的矛盾.
\[W(next) = f(W(now),congestion signals, K);\] \[R(next) = g(R(now), congestion signals, K):\]

A high value of K gives the responsiveness needed to react quickly to congestion but suffers from bad performance and instability in equilibrium when there are lags in the control loop; conversely, a low value of K can provide good performance in equilibrium but make the source sluggish during transience.

实验验证:

Design

On-Ramp设计思想非常简单: On-Ramp aims to bring down the path queuing delays as quickly as possible by pausing the flow at the sender’s end of the network when the measured OWD (which we denote as O) exceeds a threshold T.

  • Strawman solution

When OWD O exceeds T at time t, the sender immediately pauses for duration O − T.

上述做法只有在没有 feedback delay 的时候准确, 当有 feedback delay 时, the sender will actually pause for a significantly longer time than necessary.

  • Final solution

Implementation

Evaluation

Evaluation 非常细致

Thoughts

  • 这篇文章着重区分了瞬态拥塞和持久拥塞对CC的影响,但是其解决方法比较巧妙(直接用PAUSE). 还有一种想法是 能不能显式的区分出瞬态拥塞和持久拥塞,从而采用不同模式去控制 (白盒的方法)?文中认为这种思路 “fragile because it requires defining precise conditions to trigger switches between modes. Further, when a congestion control algorithm in equilibrium encounters severe congestion and drastically cuts its window or rate, it must make a difficult choice between remembering and forgetting its previous equilibrium state (rate/window). This choice is difficult because it depends on the duration of congestion, which is hard to predict. If the congestion is transient, such as an incast, the algorithm must remember its previous state so as to rapidly restore the old equilibrium without under utilization, once the incast ends”

    有待进一步研究。

参考文献

Breaking the Transience-Equilibrium Nexus: A New Approach to Datacenter Packet Transport