TCP using explicit feedback and window computation

["Saverio Mascolo" <mascolo at poliba.it>]

Dear Kalampoukas, Varma and Ramakrishnan,

I raed your interesting infocom98 paper "Explicit Window Adaptation: a
method to
enhance TCP performance" .
The paper addresses the issue of local IP nets interconnected via a high
speed ATM backbone. 
The backbone is assumed to run an efficient loss free rate control algo.
Therefore congestion and queue overflow happens at the edge IP/ATM router.
To overcome this problem you propose that the edge IP/ATM router set the
receiver's advertised window in the TCP ack to the minimum of this value
and a function of the buffer that is free at the edge IP/ATM router. 

In a report available on the web at 

http://www.cs.ucla.edu:8001/Dienst/UI/2.0/Describe/ncstrl.ucla_cs%2f980008?a

bstract=
(for an updated version, please contact me )

I am proposing  simply to set the space for the receiver's advertised
window in the TCP ack to the minimum of the free buffer encountered along
 the path.
You propose to set the window to

Window=min(Advertised Window,f(Bf)) 

Regarding the function of free buffer f(Bf),you say that "a difficult task
in such an algo is to design the feedback function f(Bf)". You propose a
simple slow time varying adaptive
proportional feedback 

f(Bf)=k*(B-Q(t))=k(Bf)

that is, the window is proportional through k to the free space B-Q(t),
where B
is buffer capacity, Q queue level. 
k is slowly adapted to take into accounts traffic condition, round trip
time etc. 
You propose an additive increasing and multiplicative decreasing algo to
tune k.
This part of the algo is not rigorous and basically follows the usual
heuristic arguments and need some "artistic" parameter tuning. Therefore,
full link utilization cannot be ensured in all realistic network scenarios.

On the other hand, the algorithm I am proposing with Prof. Mario Gerla
(Ucla) computes the control window
using a rigorous control theory based approach .
In this way the algorithm works well in any network scenario. There is no
tuning of parameter using heuristic arguments. There is no computation of
the window f(Bf) that must be performed at the network nodes.

Basically the only change we propose is to set the space reserved in the
TCP ack for the receiver's advertised window to the minimum of the free
buffers encountered along the path. 
The nice thing that is worth noting and really wonder me is that the
current
TCP already implements Smith's principle (i.e. a rigorous control theory
for time-delay systems) even if it does not know it and it does not
implement it completely! In other words, the algorithm we designed using 
classical control theory ( the first time it was developed for ABR traffic)
finally end up in a control law that is a direct generalization of the
current TCP. Therefore, the current TCP can be easily modified allowing
routers to stamp minimum free buffer encountered along the path-flow in the
TCP ack.

Saverio Mascolo
---------------------------------------------------------------------
Ph.d  Saverio Mascolo
Assistant Professor
Dipartimento di Elettrotecnica ed Elettronica, Politecnico di Bari
Via Orabona 4, 70125 Bari, ITALY

Email:mascolo at poliba.it
Phone:+39 80 5460621
Fax: +39 80 5460410




==============================================
REPORT Abstract
TCP CONGESTION AVOIDANCE USING EXPLICIT
BUFFER NOTIFICATION

Saverio Mascolo and Mario Gerla
980008
February 1, 1998

This paper proposes to modify the TCP congestion control mechanism by
allowing network
routers to supply Explicit Buffer Notification (EBN) to the TCP data
sender. The TCP sender
runs an algorithm that uses the EBN feedback to compute the control window.
The EBN plays
the role of a generalized Advertised Window so that the EBN TCP results to
be a natural
extension of the current TCP flow control. The key feature of the proposed
congestion avoidance
algorithm is that it is not based on packet loss to detect congestion. As a
consequence, it results
to be suitable for applications that are sensitive to the delay or loss of
individual packets, such as
interactive traffic, and for data transfer over wireless links where packet
loss will occur more
often due to unreliable links than due to congestion. The proposed control
works over connection
paths with any bandwidth(delay product. Its strength lies in the fact that
it is designed via
classical control theory and, therefore, its dynamic properties, such as no
cell loss and full link
utilization, can be rigorously and easily shown via mathematical analysis.
Therefore the validation is general and not restricted to the simulated
network and traffic scenario. Finally, Fair Queuing
at the routers decouples the flows and ensures fairness.

---------------------------------------------------------------------
Dr. Eng. Saverio Mascolo
Assistant Professor
Dipartimento di Elettrotecnica ed Elettronica
Politecnico di Bari
Via Orabona 4
70125 Bari-ITALY

Email:mascolo at poliba.it
Phone:+39 80 5460621
Fax: +39 80 5460410


---------------------------------------------------------------------
Dr. Eng. Saverio Mascolo
Assistant Professor
Dipartimento di Elettrotecnica ed Elettronica
Politecnico di Bari
Via Orabona 4
70125 Bari-ITALY

Email:mascolo at poliba.it
Phone:+39 80 5460621
Fax: +39 80 5460410
Web page http://www-dee.poliba.it