Exploring InfiniBand Congestion Control

InfiniBand is a state-of-the-art high performance networking technology used in many of the worlds largest supercomputers, clouds and modern datacenters where high bandwidth and low latency is required. InfiniBand supports congestion control with a parameterised mechanism, but proper guidelines or policies for using congestion control in InfiniBand do not exists.

Comparison of simulation results and hardware results for InfiniBand congestion control.

The InfiniBand standard describes a congestion control mechanism for detecting and resolving congestion.  This mechanism is rich in the way that it specifies a set of parameters that can be tuned in order to achieve effective congestion control and the exact behaviour of mechanism depends upon the values of these parameters. There is, however, limited experience on how to configure these parameters. The purpose of the thesis is to study a subset of these parameters in order to understand their impact on the behaviour of the mechanism and evaluate the performance gains.

What you will do:
  • You will configure and tune congestion control using the cutting-edge research infrastructure eX3 and through simulations with OMNet++
  • You will evaluate the performance gains of congestion control.
  • You will analyse the parameter space of several of the InfiniBand congestion control parameters through a parameter study.
  • Potentially publish your findings at a scientific venue!
What you will learn:
  • You will get to know the InfiniBand architecture, a state of the art high performance networking technology.
  • You will obtain a thorough understanding of congestion control in general and its specific implementation in InfiniBand.
  • You will learn about performance analysis and parameter studies using mesasurements and simulations.
For more information, please contact Ernst Gunnar Gran (ernstgr@simula.no)
Emneord: Congestion Control, Head-of-line Blocking, InfiniBand, Interconnection Networks
Publisert 25. okt. 2021 11:24 - Sist endret 25. okt. 2021 11:24