Cell Relay Archive

["Johannes Krohn" <KROHN@ffm.pandacom.de>]

It@s in a way the old story of comparing apples with  pears. ABR is a 
service category defined to use up bandwith not used by other service 
categories (CBR, rt-VBR, nrt-VBR). To implement this service category 
the atm forum published the TM 4.0 spec describing a framework of 
actions to be taken in end systems and network elements to implement 
a ABR service.

PPD and EPD are actions that can be taken by a network element 
(switch) in reaction to beeing in a congested state. In contrast to 
dropping random cells of random connections on a congested link the 
switch instead drops all cells based on adaption layer information. 
For PPD this means that all cells excluding the EOM cell on one VCC 
are dropped or in the case of EPD all cells from the first cell after 
an EOM cell up to the next EOM cell resulting in a damaged AAL-5 PDU 
in the case of PPD and in a completely dropped AAL-5 PDU in the case 
of EPD.

Where both concepts come together is in comparing the QoS of data 
services on ABR with the QoS of data services on UBR with or without 
PPD/EPD. Some tests show that UBR with PPD/EPD is comparable to ABR. 
When comparing both one should take into account the way the switches 
or end systems implement the ABR service. The TM 4.0 framework 
describes several actions a switch could take in reaction to 
congestion:
   - sending backward RM cells (BRM)
   - setting the congestion indication in forwarded data cells (EFCI)
   - modifying the ER (explicit rate) in FRM and BRM cells.
   - implementing virtual source/destination (VS/VD) behaviour by 
     terminating the FRM and BRM flows.
The choice the switch takes out of these actions should influence the 
QoS of the ABR service category. In addition to this it gets more 
complex in the context of large networks (not only one switch) where 
every switches actions add to the overall performance of a link. I 
would expect in large networks switches with VS/VD behaviour to 
outperform switches only implementing EFCI and ER modification 
because the length of the feedback loop and thus the reaction time of 
the traffic source on the feedback information is reduced leading to 
smaller cell buffer sizes needed to handle the traffic.

In my opinion it is no question of TM4.0 or EPD/PPD because even ABR 
service implemented in conformance to TM4.0 could gain from EPD/PPD. 
This is because TM4.0 can not guarantee zero cell loss to an ABR 
connection, but only tries to _minimize_ cell loss for compliant ABR 
connections. This means that even a switch implementing TM4.0 can 
come to a congested state and thus has to drop cells and here it is
definitely is better to do EPD/PPD on non compliant or compliant ABR 
connections than to simply drop random cells.

Another point is to analyze the relationship between ABR flow control 
and TCP flow control on the network layer when measuring TCP 
performance over ABR. During the 39th IETF meeting in Munich there 
was an interesting presentation on the interaction of the rate based 
ABR flow control and the credit based TCP flow control indicating 
that an end system could gain some quality out of combining TCP level 
flow control and ATM level flow control.

There is by the way another framework for implementing the ABR 
service category based on credits (quantum flow control; 
http://www.qfc.org) 

Johannes
---------------------------------------------------------------------------
Pan Dacom GmbH
Training/Consulting
Johannes Krohn
Robert-Bosch-Str. 32
D-63303 Dreieich

Tel.: +49-6103/932-137
Fax: +49-6103/932-400
Email: krohn@ffm.pandacom.de