ABR - Available Bit Rate

The bit rate left after the predictive and guaranteed service traffic
(CBR/VBR) is served. In essence, it is simply a fair share of the
remaining bandwidth amongst the VPs and VCs that have asked for this
service. 

Attention is being focused on the use of closed-loop controls to avoid or
significantly limit cell loss due to congestion for the ABR and VBR+
services.  Two camps formed within the ATM Forum. One centered around
rate-based controls which use a FECN/BECN like mechanism, and the other
around a per-VC flow control mechanism which uses a buffer-credit scheme.

In October 1994, the Forum decided on a rate-based approach to the 
congestion control problem.  Work is now proceeding on the details of 
providing a solution using this approach.

FCVC (flow-controlled virtual circuits), having been rejected by the ATM
Forum, are described here for your historical edification. FCVCs keep
count of the occupied buffers for each circuit at the downstream hop.  At
the upstream end, the switch (or terminal) stops transmitting on an ABR
circuit if its downstream buffer occupancy reaches a per-VC limit. At the
downstream end, the switch (or terminal) sends back a "credit" (perhaps
encoded as a next-cell-to-forward sequence number) after forwarding some
number of cells on each circuit. 

If you send back a credit record for roughly every other cell forwarded on
each VC, and you pack 5 credit records into one credit cell, then the
reverse credit bandwidth is one tenth of the forward data bandwidth. The
implementors consider one tenth of link bandwidth to be a reasonable upper
bound on control traffic.  By changing the credit frequency parameter, you
can reduce the control traffic bandwidth as you please. 

Another technique is to piggy-back a credit message on every ATM cell sent
on the link (and to piggy-back on an unassigned cell if there's no real
cell to send at the moment).  To do this, the two ends of an ATM link
agree via the ILMI MIB to mask off unneeded bits in the ATM cell header,
and to squeeze a credit message into those bits.  This approach is much
cheaper than batching credits.  But it limits the UNI to 2^14 VCIs with no
VPIs or GFC field (2^12 VCIs with GFC but no VPI), which is fine for LAN
applications.  In the absence of a standard protocol for FCVCs, Digital's
initial LAN equipment uses this latter method, while WAN equipment will
use a batched-credit encoding. (Recall that FCVCs use entirely hop-by-hop
control procedures, so every ATM switch also serves as a flow control
protocol gateway. Interoperation of different wire protocols for credits
is therefore an inherent feature of FCVC-capable ATM switches. 
Interoperation of end-to-end rate control procedures from different
vendors is a much harder problem.)

Variations within each of the above approachs continue to emerge.

Further Reading:

H.T. Kung, Robert Morrid, Thomas Charuhas, Dong Lin:
"Use of Link-by-Link Flow Control in Maximizing ATM Network
Performance: Simulation Results",
Proc. IEEE Hot Interconnects Symposium '93,
Palo Alto, CA, 5/6 August 1993

The above definition was gleaned from posts on the comp.dcom.cell-relay
newsgroup from several individuals including Jon Crowcroft (University
College - London), Greg Waters (Digital Equipment Corp) and Christoph
Bernhardt (Institut EURECOM, France).