-----Original Message-----
From: Fred Baker [SMTP:fred(_at_)CISCO(_dot_)COM]
Sent: Friday, March 10, 2000 5:32 AM
To: Jianbo Huang
Cc: ietf(_at_)ietf(_dot_)org
Subject: Re: Critically compare the congestion control on TCP/IP
and ATM?
At 05:57 PM 3/9/00 +0800, Jianbo Huang wrote:
Dear Sirs and Madams,
A friend of mine are working on the paper on "critically compare the
congestion control on TCP/IP and ATM", and she ask me for help. But I
do
not get much idea on the "congestion control on ATM". So, is there
anyone
can give me any idea on this topic, while my friend and I processing
on this?
ATM has a carefully defined traffic management architecture. Read:
www.atmforum.com/pub/approved-specs/af-tm-0121.000.pdf
In short, it supports a strong distinction between elastic and inelastic
traffic, and provides distinct services to support each.
For inelastic traffic, there is an open loop control system which shapes and
polices
traffic to traffic contracts defined by token buckets. Signalling triggers
admission
control and bandwidth reservation. Meaningful guarantees on delay, delay
jitter and
loss can be offered to traffic streams that conform to the traffic contract.
For elastic traffic there are several options:
An available bitrate (ABR) service category provides a control loop for
matching the rate
of each virtual connection (VC) to the larger of a nominal fair share of the
bottleneck bandwidth of the path, and a negotiated minimum. End-system
behavior is standardized. Switches can either directly indicate the allowed
cell rate in resource management cells which are sent periodically, or can use
a binary feedback system (mostly for backward compatibility). ABR provides
excellent isolation, fairness, and protection against misbehaved VCs. When
the end-system behavior rules are honored, there is little if any cell loss.
ABR is also resistant to performance problems associated with errored links,
highly asymetrical links and long bandwidth-delay product links.
The unspecified bitrate (UBR) and guaranteed frame rate (GFR) service
categories offer a best-effort service like IP. They depend on packet discard
and end-to-end behavior (as in TCP) for stability. Modern implementations
have a separate queue for each VC, and provide at least that level of
isolation and fairness. There is no admission control.
that's easy; there isn't any.
See the ATM Forum Traffic Managment specification
There is ingress port policing, which is
something different,
No, it's one component of the system, especially as applied to inelastic
traffic.
and there may be PNNI call routing.
That's another component of the system, but operates on a different timescale,
and only in the context of signalling and admission control (as well as being
a critical network optimization)
But there is
not
anything that corresponds to what TCP expects from its underlying
layers.
What do you think is missing for UBR and GFR? The semantics are effectively
the same: feedback control using packet loss as a signal.
There have been some papers written and a fair bit of experience with
a
technique for mitigating this, called Early packet Discard. In
essence, if
a link is becoming congested, rather than dropping a single cell, if
it has
to drop an AAL5 cell it drops the entire packet containing the cell.
This
may sound odd, but it is actually quite sensible - if the other cells
were
not also dropped, then they would uselessly occupy bandwidth on
subsequent
links, and at the final delivery point would consume memory
unnecessarily
until the SAR was able to determine that the cell had been dropped.
Ah, the segmentation and reassembly problem. EPD is not a congestion control
mechanism per se, but rather a correction to an impedence mismatch between the
minimum unit that can be dropped (a cell) and the minimum unit which is useful
(a packet). There were a couple of papers about this in 1993 (see Sally
Floyd's web page), and it was incorporated into the ATM standards. All modern
implementations support EPD and PPD. Certainly not a point that I would pick
as being one of the most salient in a discussion of ATM congestion control.