Overview
Microburst triggered congestion affects environments
where unsynchronised data sources send data via a common
path, which converges on the wire as a highly synchronised
and bursty traffic spike. Core networking equipment is
generally provisioned to adequately handle full line rate
traffic from the network's edges and can therefore cope
with such bursts. As the burst moves closer to the
destination, the potential for queue build up and end host
processing overhead to cause momentary congestion
increases.
Bursts directed at the same end point are the most
problematic, and this scenario is commonly referred to as
the incast problem. Typically, the switch port closest to
the end host becomes inundated with packets, its buffer
overflows, and the tail of the burst is dropped at the
switch. The faster the data rate and lower the latency,
the more packets that can potentially be lost.
Clustered data centre environments with low latency,
high bandwidth connectivity are the obvious places where
microbursts and incast can be observed. Other scenarios found
commonly on the wider Internet also have the potential to
experience such problems and require examination as well.
Project Goals
- Analyse anonymised network trace
information obtained from Google's production networks,
looking for existence and details related to the nature
of congestion experienced.
- Run evaluative and exploratory
testing of congestion management schemes based on real
world scenarios/conditions identified by the Google
production network trace analysis.
- Provide new insights and perhaps
make recommendations to both Google and the Internet
community at large based on the findings and outcomes of
the project.
Program Members
Collaborators
This research draws
on anonymised data provided by Google Inc., to
promote a greater common understanding of the
web.
This project has concluded, and in 2013/2014 the research topic is being pursued here.