From: Lillie Kittredge (kittredl@u.washington.edu)
Date: Sun Oct 17 2004 - 23:23:31 PDT
RED gateways
This paper shows the evolution of the response to the problem of
congestion control. It discusses Random Early Detection gateways for
congestion avoidance.
The previous paper for this week postulated that congestion control might
effectively be implemented in gateways; this paper shows the realization
of that suggestion. The authors improve the robustness of earlier
congestion control by allowing their RED gateways to work with either
protocols in which a dropped packet or a flag in a header indicates "slow
down! too much!". This demonstrates the layering idea we've been talking
about: the congestion control is now independent of the transport layer.
They also improve performance by detecting congestion early on, keeping
track of queue lengths and acting when the average length is too long.
The fact that they use average length instead of responding to any single
long queue means that they don't punish bursty traffic the way other
congestion control measures do.
The other main thing about this is that, when some but not all of the
flows need to slow down, they mark or drop packets probabilistically.
This adds some randomness that keeps nodes from getting synchronized, and
having to cut back their flows and go through slow-start at the same time.
Lastly, this takes into account factors such as whether the nodes in the
network might not be cooperating. Though the gateways don't actually
punish misbehaving users, they can report the misbehavior to a higher
level. The fact that this paper even considers the possibility of naughty
users, something ignored in the previous paper this week, indicates the
changes in networking between '88 and '93.
This archive was generated by hypermail 2.1.6 : Sun Oct 17 2004 - 23:23:31 PDT