NEURAL NETWORK BASED FLOW CONTROL
FOR ABR SERVICES IN ATM NETWORKS
NEURAL NETWORK BASED FLOW CONTROL FOR ABR
SERVICES IN ATM NETWORKS
ABSTRACT information such as data, audio and
This paper deals with a flow video. Different types of user traffic
control scheme for multicast ABR have different requirement on
services in ATM networks. The goal bandwidth, loss ratio and delay
of this paper is to desin an optimal which are characterized by a set of
rate control to deal with the parameters. Based on these traffic
variation in RM-Cell round trip time parameters, the ATM network setup
(RTT) resulting from dynamic drift connection (or VC-Virtual circuit)
of the bottleneck in a multicast tree. from a source to destination. A
For a given finite buffer capacity the connection runs through a series of
control scheme maximizes average intermediate switch nodes, where it
throughput and bounds end-to-end shares link bandwidth and buffer
delay. By applying Neural Network space with other connections. Thus
based rate control, the proposed the traffic rate flowing through a
scheme not only makes the rate switch depends on the number of
process converge to the connections and the source rates of
neighborhood of link bandwidth but these connections become active
also confines the queue length simultaneously, or some connections
fluctuation to a regime bounded by a increases their rates unlimitedly,
buffer capacity (thus guaranteeing queues buildup at bottlenecked
loss less transmission). switches. Eventually, the buffer
capacity is exceeded and cells are
Introduction
dropped, resulting in low
An ATM network can throughput, a large delay, and even
transport a wide varity of network blockage. To prevent a
network from failing into this kind daunting task. (Multicase includes
of congestion, an efficient flow point-to-multipoint, Multipoint-to-
control is required. point and Multipoint-to-Multipoint
transmission. However for the
Available Bit Rate (ABR) convenience of presentation, in this
services, which is suitable for paper use the narrow – sense
various data communications can definition for multicast, which
maximize network bandwidth stands for the point-to-multipoint
utilization and avoid congestion. It transmission).
provides better service for data
traffic by periodically advising
sources about the rates at which they
should be transmitting. The
switches monitor their load and Fig. Network Model with A
divide the available bandwidth Single Bottleneck Link
fairly among active flows. This
The Closed Loop Rate Control
allows competing sources to get a
Mechanism
fair share of the bandwidth from the
ATM Forum defines the
switches to the sources is indicated
behavior of the Source End system
in RM cells, which are periodically
(SEM) and the Destination End
generated by the sources and turned
System (DES).
around by the destinations.
Denote a RTT by , the
An ABR flow-control
propagation delay between the SES
algorithm consists of two
and ATM switch by 1, and between
components: determining the
ATM switch and the DES, i.e. the
bottleneck link bandwidth and
propagation delay on the bottleneck
buffer capacity. In a multicast ABR
link by 2. Thus = 2 (1+2) if the
connections, determine the
buffer length at the ATM switch
bottleneck link bandwidth is a
exceeds Q H, the switch detects back from the destination to the
congestion. Congestion is regarded source, switches (Fig) process RM
to be terminated of the buffer length cells. When an RM cell arrives at the
drops below QL. destination, the destination changes
the direction bit (DIR) in the cell and
ABR Flow Control Mechanism
returns it to the source are called
In ATM networks, ABR traffic
Backward Resource Management
sources adjust their transmission
(BRM) cells, BRM cells bring
rates dynamically between a pre-
updated network state information
specified Minimum Cell Rate (MCR)
to the sources.
and Peak Cell rate, based on the
amount of network bandwidth left
unused by higher priority classes.
The rate adjustment is done using a
closed-loop feedback mechanism,
using RM sells. RM cells convey
Fig. ABR Traffic Management
control information to ABR traffic
Model
flows about the state of the network,
In Explicit-Rate (ER) based
such as congestion state and
congestion control schemes BRM
bandwidth availability.
cells tell the source exactly what
The ABR flow control
transmission rate to use for outgoing
mechanism is called closed-loop
traffic. This information is contained
since it uses feedback information
in the ER field of the RM cell. The
from the network to control the rate
rate at which a source is allowed to
of each source. Forward Resource
transmit cells is called the Allowed
Management (FRM) cells are
Cell Rate(ACR). The ACR is initially
generated by sources and inserted
set to a default value called the
into the outgoing data cell stream.
Initial Cell Rate (ICR). It is always
On their way to the destination and
between the MCR and the PCR. The another switch(or destination), and
source puts its ACR value into likely cause transmitted cells to be
Current Cell Rate (CCR) field of lost at bottleneck points.
outgoing RM cells, while the rate at
When a source receives a
which it wishes to transmit cells is
BRM cell. It computes its allowed
put into the ER field. RM cells are
cell rate (ACR) using information
generated by the source after
from CCR and ER field, and other
every(NRM-1) data cells are
information from the BRM cell.
transmitted, where NRM is a
parameter to the ABR traffic control ERICA Algorithm
algorithm. When an RM cell arrives The ERICA (Explicit Rate
at the destination , if the destination Indication for Congestion
is congested and cannot support the Avoidance) algorithm tries to
rate in the ER field, the destination achieve a fair and efficient allocation
reduces the ER to whatever rate it of the available bandwidth to
can support. The returning BRM cell contending sources. The basic idea
will convey this information to the in ERICA is to monitor, at each
source. switch, the incoming cell rates of
each ABR traffic source, and
As an RM cell travels back
compare the aggregate ABR traffic
through the network, each switch
demand to the desired target
examines the cell to determine if it
utilization U for ABR traffic sources(
can support the ER rate for the
typically U=0.95% of available ABR
requested connection. If the ER is too
capacity in LAN environment, and
high for a switch, the switch reduces
U=0.90% of available ABR capacity
this value to a rate than it can
in WAN environment). If the
support. Note that no switch is
aggregate demand exceeds the target
allowed to increase the ER, because
load, then traffic source rates can be
doing so violate the rate set by
increased. If the aggregate demand
exceeds the target load, then traffic Buffer Congestion: If the
source rates must be decreased. A maximum queue length Q max and a
parameter (e.g., =0.2) is used to switch exceeds the target buffer
determine what constitutes an occupancy Q goal, where QH 0 and 0 0; consolidated RM-cell is sent
(iii) If LBCI=1 and BCI=0, AIR is upward form the switch. Resp-
increased multiplicatively by branch-vec(i) is set to 1 if an
the same factor of q. For all feedback RM-cell is received
these three cases, the rate- from the I-th downstream
decrease parameter branch. The connection pattern of
conn-patt-vec each time when the randomness of network
non-recponsive branch is environments, and RM-cell
detected or a new connection processing and queuing delays,
request is received from a instantaneous variations of
downstream branch. bottleneck bandwidths, which are
Simulation Results very difficult to deal with
Using the network Simulator analytically.
2(NS2), conducted extensive
simulations for concurrent multiple
multicast VCs with multiple Fig. Simulation Model for Multiple
bottlenecks to study the multicast Ves
performance of the proposed scheme
with Neural-control, and compare it Simulaion Model
with schemes without Neural- The simulated network is
control. By removing the shown in fig which consists of three
assumptions made for the modeling multicast
analysis, the simulation experiments VCS running through four switches
accurately capture the dynamics of SW1,…,SW4 connected by three
real networks, such as the noise- links L1,L2,L3. Si is the source of Vci
effect of RM-cell RTT due to the I=1, 2, 3 and Rij is Si’s jth receiver. So
VC2 and VC3 share L1 and L3 simulation remain the same as those
respectively, with VC1.S1 is a used in the analytical solutions for
president ABR source which comparison purposes. Specifically,
generates the main data traffic flow. Qn=50 cells, Qgoal=400 cells, =0.4
S2 and S3 are two periodic on-off ms, q=0.6, p=16.67 cells/ms2, and
ABR sources with on-period=360 ms Ro=30 cells/ms; VC1’s 0=22.9 cells
and off-period =1011 ms, ms2. Let S1 start at t=0, s2 at t=160
respectively, which mimic cross- ms, and S3 at t=822 ms such that S2
traffic noises, causing the bandwidth and S3 generate the cross-
to vary dynamically at the
Fig. Dynamic performance of
bottleneck bandwidth with Neural
Network Control
traffic noises against the main data
Fig. Dynamic Performance of
traffic flow at the potential
Bottleneck Bandwidth without
Neural Network Control bottlenecks L1 and L3 with the
respective on-periods appearing
bottlenecks. Set Li’s bandwidth
alternatively without any overlap in
capacity i to (l) 1=3=155.52 Mb/s
time. Results of the simulation are
and (2) 2=300Mb/s, forcing the
given in figure.
potential bottlenecks L1 and L3 to
show up. Letting all link’s delays be Conclusion
1 ms s1’s RM-cell RTTs via R16,R17,R18 In this paper it is proposed
equal 4ms which is 2 times of S1’s and analyzed a flow control scheme
RM-cell RTTs via R11,R12,R13. the for ATM ABR multicast services,
flow control parameters used in the which scales well and efficient in
dealing with the variations in the
multicast –tree structure and RM –
cell RTT. Also developed the Neural
Network based control algorithm to
handle the variation of RM-cell RTT.
The simulation experiments verify
the proposed scheme to the other
schemes in dealing with the
variations of RM-cell RTT and link
bandwidth, and achieving fairness
in both buffer and bandwidth
occupancies.