Sie sind auf Seite 1von 6

An Enhancement of Multicast Congestion Control

over Hybrid Wired/Wireless Networks


*


Yan Liu

, Cheng Peng Fu

, Zong Kai Yang

, Bu Sung Lee

Huazhong University of Science and Technology, China


Nanyang Technological University, Singapore




Abstract In wireless networks, random loss due to bit error
leads to significant performance degradation to conventional
multicast congestion control schemes. In this paper, we propose
and study an enhanced multicast congestion control scheme,
called EPGMCC (Enhanced PGMCC), which achieves significant
performance improvement over PGMCC over loss-prone wireless
links. The key idea of EPGMCC is to discriminate random loss
from congestion loss according to the measured network conges-
tion level, and perform different actions accordingly. Specifically:
1) at the sender, it adapts the AIMD algorithms in TCP Veno [3]
to avoid the unnecessary window-halving induced by random loss
and 2) at the receiver, a new scheme is designed to better measure
the loss rate and thus help the sender to select the right
representative of the group. Our extensive simulation results
demonstrate that EPGMCC achieves significant throughput
improvement over PGMCC in hybrid wired/wireless networks,
furthermore, the improved throughput achieved by EPGMCC is
not grabbed from the bandwidth of coexisting connections but
from the utilization of available bandwidth that left unused.
Throughput improvement of up to 40% can be demonstrated
over typical wireless access link with 1% random loss rate.
Keywords congestion control; multicast; wireless network;
random loss;congestion loss.

I. INTRODUCTION
Multicast is an efficient technology that transmitting data
from sender(s) to a group of receivers. It enables many
applications ranging from data dissemination (where reliability
is required) to videoconference (where reliability is not
necessary). With the increasing popularity of multicast
applications, the wide deployment of multicast transport
protocols would cause congestion collapse if they cannot
provide adequate congestion control mechanisms. The IETF
reliable multicast criterion [10] requires that each multicast
transport protocol proposal to include an analysis of whether
the protocol has congestion control mechanisms strong enough
to cope with deployment in current heterogeneous Internet,
where wireless links, ADSL links and satellites links are widely
deployed.
In the past decade, multicast congestion control (MCC) has
been one of the most important and active R&D areas, and
many schemes [1], [2], [6]-[8] have been proposed to meet the
diverse requirements of different applications. Pragmatic
general multicast congestion control (PGMCC) [1] is one of the
most promising MCC schemes, and is presented and drafted by
IETF Reliable Multicast Transport (rmt) working group [10]. It
is a single-rate, window-based congestion control scheme, and
achieves TCP-friendliness, scalability, stability and fast
response to variations in network conditions.
However, these MCC schemes are designed based on the
assumption that congestion is the main reason for packet loss,
which cannot be applied to wireless environment. It is well
known that wireless technologies are experiencing a dramatic
increasing development, and wireless devices, in the form of
wireless local area networks (WLAN), cellular networks and
etc., are widely deployed and playing a more and more
important role in interconnecting between the end users and
backbone networks. Similar to the traditional TCP suffering
from lossy wireless links, the conventional multicast
congestion control schemes experience great performance
degradation for misinterpreting the packet loss due to bit error
as those due to congestion and decreasing the data rate
accordingly.
In this paper, we study the behavior of PGMCC running
over hybrid wired/wireless networks, and show that the trivial
random loss (average random loss rate is vary between 10
-2
and
10
-8
as conditions change in wireless environment [12]) can
lead to great performance degradation to PGMCC connections.
Based on this observation, an Enhanced PGMCC (namely
EPGMCC below) scheme is proposed to alleviate this
performance degradation. The proposed scheme is based on
two separate but complementary mechanisms: 1) at the sender,
the AIMD (Additive Increase and Multiplicative Decrease)
algorithms in TCP Veno [3] are adapted to refine the window-
based controller in PGMCC, and thus avoid the unnecessary
window-halving; 2) at the receiver, a new metric of forward-
path delay is introduced to distinguish between the congestion
loss and random loss, so to measure the loss rate much accurate
and help the sender to select the right representative of the
group. Our extensive simulation results demonstrate that
EPGMCC achieves significant performance improvement over
PGMCC over loss-prone wireless links, while keeps the TCP-
friendliness feature of PGMCC in wired networks.
Furthermore, EPGMCC coexists with TCP and PGMCC
harmlessly, for the improved throughput achieved by
EPGMCC is not robbed from the bandwidth of coexisting
connections but from the utilization of the available bandwidth
that left unused. EPGMCC is compatible with TCP.
The paper is structured as follows. Section II presents the
PGMCC mechanism briefly. In Section III, we study the
performance degradation of PGMCC over wireless access links
*This work was supported by the National Science Foundation of China
(NSFC) under grant number 60202005.
WCNC 2004 / IEEE Communications Society 2563 0-7803-8344-3/04/$20.00 2004 IEEE
and describe EPGMCCs mechanism in detail. The
performance of EPGMCC is evaluated in Section IV, and we
draw conclusions and present future work in Section V.
II. PGMCC MECHANISM
PGMCC is a single rate window-based congestion control
scheme, and adjusts its sending rate according to the slowest
receiver (namely the ACKER). It uses a tight control loop
which closely mimics TCP behavior between the sender and
the ACKER, and is suitable for both reliable and non-reliable
applications. For missing data packets are not necessarily
retransmitted in PGMCC, the window-based control loop in
TCP is simulated using a token-based mechanism that
decouples the congestion control scheme from retransmission
state. In this token-based mechanism, two state variables are
maintained:
A window size W, which denotes the current window
size in packets;
A token count T, which is used to regulate the
transmission of data packets.
The parameters W and T are updated upon receiving
feedbacks from ACKER according to the following rules:
INITIALIZATION W = T = 1
ACK RECEPTION, NO LOSS DETECTED
if (W < SSTHRESH) then
//slow start
D = min(N, SSTHRESH - W)
N = N D
T = T + 2*D
W = W + D
endif

T = T + N * ( 1 + 1/W )
W = W + N/W
N is the number of new packets acknowledged by the
incoming ACK.
ACK RECEPTION, LOSS DETECTED
T = T - W/2 , W = W/2
PACKET TRANSMISSION
T = T - 1
The ACKER is responsible for providing timely feedback
just in the same way as TCP receiver. For ACKER is selected
as the receiver experiencing the worst networks conditions, it
may change over time since network traffic varies and
receivers are continuously joining and leaving the group.
PGMCC uses a fast and low-overhead procedure to select the
ACKER dynamically: each receiver measures the loss rate and
feedbacks this information to the sender either in ACK (for
ACKER only) or NAK (for all receivers) messages; the sender
uses this feedback to measure the RTT to each receiver; both
loss rate and RTT are fed into a TCP throughput equation to
determine the expected throughput to that receiver; the receiver
with the lowest expected throughput is selected to act as the
ACKER.
III. PERFORMANCE DEGRADATION OF PGMCC OVER
WIRELESS LINKS AND ITS ENHANCEMENT
A. Performance Study
In this subsection, we study the performance degradation of
PGMCC over loss-prone wireless links, and show that the
random loss due to bit error but not due to congestion
influences the performance of PGMCC significantly.







Fig. 1. Network scenario

The network scenario considered in this subsection is
shown in Fig. 1. The sender and receivers of PGMCC session
are labeled as S and R
i
, and L1 is the bottleneck link with
capacity of 1Mbit/s, propagation delay of 50ms and queue limit
of 12. Receivers access the network through two types of links:
one type is wired access link with capacity of 10Mbit/s and
propagation delay of 0.1ms; another type is wireless access link
with capacity of 5Mbit/s, propagation delay of 0.1ms and
random loss rate of 1%.
In conventional congestion control schemes, the packet loss
is regarded as the sign of the network congestion, and the
window size is halved upon loss detected. This assumption
works well in wired networks, but is brings unnecessarily
decreasing of the window size over lossy wireless links, and
thus degrades the throughput of the sessions dramatically.
Consider running PGMCC over hybrid wired/wireless
networks shown in Fig. 1, according to the ACKER selection
mechanism in PGMCC, the receiver accesses through wireless
link will be selected as the ACKER for it has higher loss rate
that induced by random loss. Upon selecting the receiver
accesses through wireless links as the ACKER, the session will
experience great performance degradation for packet losses due
to random loss evoke the congestion avoidance mechanism in
PGMCC.
For illustration, Fig. 2(a) and (b) show the different beha-
vior between the ACKER accesses through wired links without
random loss and though wireless links with random loss rate
1%.
As seen in Fig. 2(a), the halving of the congestion window
occurs much frequently when the receiver accesses through
wireless links is selected as the ACKER. This unnecessary
window halving results in the under utilization of the available
bandwidth, as seen in Fig. 2(b). The ACKER can receive about
36.0k packets when its access links are wired, but its only
15.0k packets as the access links are wireless. From the figure,
we can see that even a trivial random loss can result in great
performance degradation to PGMCC connections over wireless
links.
R1
R2
Rn
S
L1

WCNC 2004 / IEEE Communications Society 2564 0-7803-8344-3/04/$20.00 2004 IEEE


(a) PGMCC congestion window (b) PGMCC sequence number
Fig. 2. PGMCC behavior
B. Enhanced Scheme
As in unicast, distinguishing between congestion loss and
random loss effectively and performing separate actions to deal
with them are fundamental ideas to improve the performance of
multicast congestion control mechanism over hybrid
wired/wireless networks. Compared with the work on unicast
[3][9], to improve the throughput of multicast congestion
control mechanism over wireless links is much complicated
inherently. Our enhancement includes two main mechanisms:
A refined window-based controller
This window-based controller is simply an adaptation
of the TCP Veno [3], which is proved to be an efficient
and robust enhancement to TCP over hybrid
wired/wireless networks.
ACKER Selection
In PGMCC, the ACKER is selected as the receiver
with the lowest expected throughput, which is computed
according to the measured loss rate and RTT. In current
loss rate measurement scheme, there is no mechanism to
distinguish between the congestion loss and random loss,
so the measured loss rate reflects the sum of the
congestion loss rate and random loss rate in hybrid
wired/wireless networks. This measurement deviates the
original idea of congestion estimation along the path. To
eliminate the negative effect brought by the random loss
on the ACKER selection, we introduce a metric of
forward-path delay to discriminate random loss from
congestion loss at the receiver, and then estimate the
congestion loss along the path correctly, thereafter, a right
ACKER is able to be selected.
B.1. Window-based controller
In PGMCC, the window-based control loop is simply an
adaptation of the conventional TCP congestion control in
which the mechanism to distinguishing between congestion
loss and random loss is not applied. In our enhancement, we
adapt the AIMD algorithms in TCP Veno [3] as refined
window-based controller.
Veno uses the measurement of N, the number of backlog at
the bottleneck queue, as an indication of the network
congestion status. N is computed as:
N =
f
* (RTT - BaseRTT)
= (W/BaseRTT W/RTT) * BaseRTT (1)
where
f
is the capacity of the bottleneck link, RTT and
BaseRTT denote the measured round-trip time and the
minimum of the measured RTT, and W is the current window
size.
In our enhancement, we adapt the idea of using N as
indication of the network congestion level and adjust the
window size accordingly. The refined window-based controller
is as follows:
The linear window opening is modified to:
if (N < )
// available bandwidth not fully utilized
T = T + (1 + 1/w),
W = W + 1/W; when each new ACK is received.
else if (N >= )
// available bandwidth fully utilized
T = T + (1 + 1/w),
W = W + 1/W. when every other new ACK is received
And the window halving is modified as:
if (N < )
//random loss is most likely to have occurred
T = T W*(1/5),
W = W * (4/5);
else // congestive loss is most likely to have occurred
T = T W/2,
W = W / 2.
EPGMCC only modifies the window additive increase and
multiplicative decrease algorithms. Other window evolutions,
such as initialization setup and slow start algorithms, are kept
intact in our enhancement.
B.2. ACKER Selection
In PGMCC, ACKER is selected according to the expected
throughput of each receiver. To compute the expected
throughput to the sender, each receiver measures the loss rate
and feeds it back to the sender. There is no packet loss
discrimination mechanism in current loss rate measurement
WCNC 2004 / IEEE Communications Society 2565 0-7803-8344-3/04/$20.00 2004 IEEE
scheme, thus the measured loss rate reflects the sum of the
congestion loss rate and random loss rate in hybrid
wired/wireless environments. This erroneous loss rate
measurement can result in erroneous selection of ACKER, and
thus degrades the performance of the whole session. To keep
the measured loss rate to be the reflection of congestion loss
rate, a new metric of forward-path delay is introduced to
discriminate random loss from congestion loss. The forward-
path delay, D_r
i
, is defined as time interval between the
packet is sent by the sender and that packet is received by the
receiver:
D_r
i
= t_r
i
t_
i
(2)
where t_r
i
and t_
i
are the time that packet i is received by
receiver r and packet i is sent by the sender. When the queue
size is zero on the forward path, the forward-path delay is
defined as BaseDelay. BaseDelay is basically the sum of the
transmission delays and propagation delays throughout the
forward path. Let the number of backlogged packets at the
queue be denoted by N, we have
D_r
i
= BaseDelay + N/
f
(3)
where
f
is the capacity of the bottleneck link on the forward
path. Accordingly, N can be estimated by
N =
f
* (D_r
i
- BaseDelay) =
f
* D_r
i
(4)
For
f
is less sensitive to instantaneous network traffic, and
it is not a trivial task to measure the
f
on each receiver, N is
estimated by the variation on forward-path delay in our
enhancement. The variation on D_r
i
, D_r
i
, is used as an
indication of whether the forward path is in congestion or not.
Specifically, if D_r
i
< when a packet loss is detected, the
receiver will assume the loss to be random loss rather than
congestion loss; otherwise, the receiver will assume the loss to
be congestion loss. is set to be the mean variation of
forward-path delay measured in last round-trip time. The loss
rate p_
i
of receiver i is computed as
p_
i
= c * p_
i-1
+ (1 - c) * x_
i
(5)
where c is a constant between 0 and 1, and
x_
i
= 0 if packet i is successfully received,
x_
i
= 0.1 if packet i is lost and D_r
i
< ,
x_
i
= 0.9 if packet i is lost and D_r
i
>= .
After the measurement of loss rate, each receiver feedbacks
this information to the sender for ACKER election according to
the procedures that defined in PGMCC.
Finally, it is worthwhile to point out the clock synchro-
nization between the sender and receiver is not required for the
measurement of the variation on forward-path delay. The clock
skew is dispelled when BaseDelay is subtracted from the
measured forward-path delay as long as clock drift is constant.
IV. PERFORMANCE EVALUATION
In this section, we investigate the behavior of EPGMCC
with the ns2 network simulator [11]. For space limitation, only
a small fraction of the conducted simulations can be presented
in this paper. In all simulations below, drop-tail queue is used
at the routers to imitate the behavior of current Internet. In
section IV-A, the general simulation scenario is described.
Fairness towards TCP and intra-protocol fairness of EPGMCC
are studied in section IV-B. Section IV-C demonstrates the
effectiveness of the network status measurement scheme. The
performance results of EPGMCC and PGMCC are compared in
lossy and non-lossy networks in section IV-D. The results
demonstrate that EPGMCC is TCP-friendly in wired networks
and achieves significant performance improvements over
PGMCC in lossy wireless networks, while co-exists with
PGMCC harmless in all circumstances.
A. Network scenario








Fig. 3. Network topology
The behavior of EPGMCC is studied using the well-known
single-bottleneck topology (shown in Fig. 3) where a number
of sessions share a common bottleneck link. The bottleneck
link has fixed capacity of , propagation delay of and queue
limit of B. A typical configuration is = 1.6Mbit/s, = 50ms
and B = 28. Two main configurations have been used for
access links of R
ij
: 1) the link with capacity of 10Mbit/s,
propagation delay of 0.1ms and no random loss and 2) the link
with the capacity of 5Mbit/s, propagation delay of 0.1ms and
random loss introduced. The former configuration mimics the
behavior of wired access links where congestion is the main
reason for packet loss, and the latter one emulates the loss-
prone wireless access links.
B. Inter- and Intra-protocol fairness
Fig. 4. Inter- and intra-protocol fairness
This subsection evaluates the inter- and intra-protocol
fairness features of EPGMCC. It is worthwhile to point out
that the inter-protocol fairness can be achieved only in wired
networks. In networks that contain wireless links, EPGMCC
outperforms PGMCC and conventional TCP as shown in next
subsection. The fairness towards TCP and intra-protocol
fairness of EPGMCC is studied by running multiple instances
R11
R1n
R2n
R21
S1
S2
Bottleneck Link

WCNC 2004 / IEEE Communications Society 2566 0-7803-8344-3/04/$20.00 2004 IEEE


(a) N vs. queue length (b) D_r
i
vs. queue length (for visibility, the plotted value
of D_r
i
is multiplied by 300)
Fig. 5. Measurement of the network congestion level
of the protocol sharing the same bottleneck link, using the
network topology in Fig. 3. Fig. 4 presents a typical simulation
result of running three connections, two using EPGMCC and
one using TCP, over non-lossy wired networks with the
bottleneck link of = 1.6Mbit/s, = 50ms and B = 28. Both
connections have the same packet size of 506Bytes. We see
that the numbers of received packets of all connections are
almost the same in Fig.4, and these three connections share the
bottleneck bandwidth with a fair manner.
In our simulation results, there is a good sharing of
bandwidth between EPGMCC and TCP connections in wired
networks with different configurations, and the fairness
between multiple EPGMCC connections can be achieved in
both wired and wireless networks.
C. Measurement of the network congestion level
To verify the measurement scheme of the network
congestion level, we have run some simulations with the
network topology in Fig. 3. The simulations were conducted by
configuring the bottleneck link with a variety of parameters and
using diverse background traffic. For observation, the
measured network congestion level was compared to the queue
length of the buffer on the bottleneck queue.
Fig. 5 shows a simulation result with the bottleneck confi-
guration of = 1.6Mbit/s, = 50ms and B = 28 and back-
ground traffic of one PGMCC connection. Fig. 5 (a) plots the
relationship between the estimated number of packet in the
bottleneck buffer measured at the sender and the actual queue
length of the bottleneck buffer. It can be seen that the measured
queue length is proportional to the actual queue length, and
tracks the change on the queue length exactly. Fig. 5 (b)
presents the relationship between the measured delay variation
D_r
i
and the queue length of the bottleneck queue (for clear
results, the plotted value of D_r
i
is multiplied by 300). From
the figure, we can see that the measured delay variation is
proportional to the actual queue length exactly.
Same results are produced in our extensive simulations. The
results demonstrate that the measured variables can track the
changes of the queue length on the bottleneck link closely both
at the sender and the receiver. For network congestion is
generally preceded by an increase on the queue length of the
bottleneck buffer, it is feasible to distinguish between the
congestion loss and random loss according to the measured
variables in EPGMCC.
D. The Performance
We have done numerous simulations to study the perfor-
mance improvement achieved by EPGMCC over PGMCC. In
this subsection, we present some typical results of two
connections sharing the common bottleneck link with the
network topology depicted in Fig.3. The bottleneck link is
configured as = 1.6Mbit/s, = 50ms and B = 28, and all
connections have the same packet size. Fig. 6 shows the
sequence number evolution of both connections.
Fig. 6(a) shows the results of two PGMCC connections, and
(b) shows the results of one EPGMCC and one PGMCC
connections over wired links with no random loss. We can see
that EPGMCC connection shares the bottleneck bandwidth
with PGMCC connection fairly over wired links.
Fig. 6(c) and (d) show the results of the same simulation
over lossy links with random loss rate 0.005. As seen in Fig.
6(d), EPGMCC connection achieves higher throughput than
coexisting PGMCC connection, and the improvement is about
30%. A noteworthy point in the figure is that, the throughput
of PGMCC connection depicted in (d) is close to that in (c). It
can be concluded that the EPGMCC connection coexists with
PGMCC harmlessly, since it achieves throughput improve-
ment by using the available bandwidth that left unused but not
grabbing from PGMCC connection.
Same conclusions can be deduced from results shown in
Fig. 6(e) and (f). As depicted in (f), the throughput of
EPGMCC connection is improved 40% over PGMCC nearly,
while the throughput of coexisting PGMCC connection is the
same as that in (e) roughly.
Our numerous simulation results show that EPGMCC
achieves significant performance improvement over PGMCC
over loss-prone wireless access links, while keeps the
friendliness to PGMCC in wired networks. Throughput
improvement of up to 40% can be demonstrated over typical
wireless access link with 1% random loss rate. Furthermore,
EPGMCC coexists with PGMCC harmlessly, for the improved
throughput achieved by EPGMCC is not grabbed from the
bandwidth of coexisting PGMCC connections but from the
utilization of the available bandwidth that left unused.
WCNC 2004 / IEEE Communications Society 2567 0-7803-8344-3/04/$20.00 2004 IEEE
(a) (c) (e)
(b) (d) (f)
Fig. 6. (a) Two PGMCC connections with no random loss. (b) One EPGMCC and one PGMCC with no random loss.
(c) Two PGMCC with random loss rate 0.005. (d) One EPGMCC and one PGMCC with random loss rate 0.005.
(e) Two PGMCC with random loss rate 0.01. (f) One EPGMCC and one PGMCC with random loss rate 0.01.
V. CONCLUSIONS AND FUTURE WORK
We have proposed an enhanced multicast congestion
control scheme, called EPGMCC, which can alleviate the
performance degradation of multicast connections over hybrid
wired/wireless networks. Our extensive simulation results
demonstrate that EPGMCC can estimate the network status
(congestion or non-congestion) closely at the sender side as
well as at the receiver side, and then determines whether the
random loss is most likely occurred or not. In wire networks,
EPGMCC keeps the TCP-friendliness feature of PGMCC. In
hybrid wired/wireless networks, a significant throughput
improvement over PGMCC is achieved by EPGMCC. As
indicated in our numerous simulations, the improved
throughput achieved by EPGMCC is not grabbed from the
bandwidth of coexisting PGMCC connections but from the
utilization of the available bandwidth that left unused.
EPGMCC is compatible with PGMCC. Throughput
improvement of up to 40% can be demonstrated over typical
wireless access link with 1% random loss rate.
The key idea of EPGMCC is to discriminate the random
loss from the congestion loss, and provide according actions to
deal with them. This idea is independent from any protocol
instances of multicast congestion control, and can be applied to
any MCC scheme to improve its performance over loss-prone
wireless links.
The accuracy of judgment to different types of packet loss
and the actions to deal with different types of packet loss play
an important role to the performance improvement and
compatibility with TCP flows, PGMCC flows and other
congestion-controlled flows. The mechanisms presented in this
paper are just instructive explore. We will further study this
direction in future work.
REFERENCES
[1] L. Rizzo, PGMCC: A TCP-friendly Single-Rate Multicast Congestion
Control Scheme, ACM SIGCOMM 00, Aug. 2000.
[2] J. Widmer and M. Handley, Extending Equation-Based Congestion
Control to Multicast Applications, ACM SIGCOMM01, San Diego,
Aug. 2001.
[3] Cheng Peng Fu, S. C. Liew, TCP Veno: TCP Enhancement for
Transmission Over Wireless Access Networks IEEE Journal of
Selected Areas in Communications, Feb. 2003
[4] Cheng Peng Fu, TCP Veno: End-to-end Congestion Control over
Heterogeneous Networrks, Ph.D dissertation, July, 2001
http://www.broadband.ie.cuhk.edu.hk.
[5] A. Mankin, A. Romanow, S. Bradner, and V. Paxson. IETF criteria for
evaluating reliable multicast transport and application protocols.
RFC2357, June 1998.
[6] J. Golestani and K. Sabnani. Fundamental observations on multicast
congestion control in the Internet. In Proceedings of IEEE
INFOCOM99, 1999.
[7] J. C. Lin, S. Paul, RMTP: A reliable multicast transport protocol, In
Proceedings of IEEE INFOCOM96, March 1996.
[8] I. Rhee, N. Ballaguru, and G. N. Rouskas. MTCP: Scalable TCP-like
congestion control for reliable multicast. In Proceedings of IEEE
INFOCOM99, 1999.
[9] Christina Parsa, J.J.Garcia-Luna-Aceves, Differentiating Congestion
vs. Random Loss: A Method for Improving TCP Performance over
Wireless Links, In Proceedings of IEEE WCNC00, 2000.
[10] http://www.ietf.org/html.charters/rmt-charter.html
[11] VINT Project, Network Simulator version 2(ns-2),
http://www.isi.edu/nsnam/ns.
[12] H. Balakrishnan and R.H. Katz, Explicit Loss Notification and
Wireless Web Performance, In Proceedings of Globecom98 (Internet
Mini Conference), November 1998.
WCNC 2004 / IEEE Communications Society 2568 0-7803-8344-3/04/$20.00 2004 IEEE

Das könnte Ihnen auch gefallen