﻿Toward Picture-perfect Streaming on the Internet ∗
Alix L.H. Chow
Computer Science Dept.
Univ. of Southern California
lhchow@cs.usc.edu
Abstract
Quality of service (QoS) in streaming of continuous media
over the Internet is poor, which is partly due to variations
in delays, bandwidth limitations, and packet losses.
Although continuous media applications can tolerate some
missing data, non-recoverable information loss degrades
these applications’ QoS. Consequently, a number of application
areas have backed away from streaming of their content
over the Internet. Inability to control the resulting visual
and auditory quality of the streamed presentation is an
important reason for such a trend.
We believe that this trend can be reversed. To this end,
this paper gives an overview of our efforts in exploring high
quality streaming through the exploitation of multiple paths
existing in the network. By high quality, we mean with significant
bandwidth requirements, of relatively long duration,
and without information loss or hiccups in data delivery.
We believe this to be a promising approach.
1 Introduction
Quality of service (QoS) in streaming of continuous media
(CM) over the Internet is poor, which is partly due
to variations in delays, bandwidth limitations, and packet
losses. Although CM applications can tolerate some missing
data, non-recoverable information loss degrades these
∗ This research has been funded in part by the NSF ANI-0070016, NSF
EIA-0091474, NSF CCR0113192, RGC, as well as the Croucher Foundation
Scholarship and the Okawa Research Award. It has also been funded
in part by the Integrated Media Systems Center, a National Science Foundation
Engineering Research Center, Cooperative Agreement No. EEC-
9529152. Any opinions, findings and conclusions or recommendations
expressed in this material are those of the author(s) and do not necessarily
reflect those of the National Science Foundation.
Leana Golubchik
CS & EE-Systems Depts, IMSC, ISI
Univ. of Southern California
leana@cs.usc.edu
John C.S. Lui
Dept. of Computer Science & Engineering
The Chinese Univ. of Hong Kong
cslui@cse.cuhk.edu.hk
Proceedings of the Second International Conference on the Quantitative Evaluation of Systems (QEST’05)
0-7695-2427-3/05 $20.00 © 2005 IEEE
applications’ QoS. Consequently, a number of application
areas (e.g., those related to the entertainment industry) have
backed away from streaming of their content over the Internet.
Inability to control the resulting visual and auditory
quality of the streamed presentation is an important reason
for such a trend.
We believe that this trend can be reversed. To this end,
this paper gives an overview of our efforts in exploring high
quality streaming through the exploitation of multiple paths
existing in the network. By high quality, we mean with
significant bandwidth requirements, of relatively long duration,
and without information loss or hiccups in data delivery.
In this paper, we will give a summary of our findings
and briefly present evidence that multi-path streaming is a
promising approach.
Our goal is to design an application-level approach, i.e.,
one that can be deployed over the Internet today, without requirements
for support from/modifications to the lower layers
of the network. This is in contrast, for instance, to an
approach which would use the IntServ model for signaling
(e.g., RSVP) and resource reservation in all routers along
the streaming path. Such an approach would suffer from
scalability and deployment problems.
Specifically, we investigate the potential benefits of providing
QoS in CM delivery through the exploitation of multiple
paths existing in the network between a set of senders
and a receiver. One advantage of this approach is that the
complexity of QoS provision can be pushed to the network
edge (an original design principle of the Internet) and
hence improve the scalability and deployment characteristics
while at the same time provide a certain QoS level.
Our focus thus far has been on providing a fundamental
understanding of the benefits of using multiple paths to
deliver CM data (such as video) destined for a particular receiver,
i.e., this data is fragmented into packets and the dif-
ferent packets take alternate routes to the receiver. This is
illustrated in the example of Figure 1, where streaming over
multiple paths is accomplished by streaming the data from
three servers, distributed over a wide area network. In this
example any server can send any fraction of the CM data;
specifically, server i sends fraction αi of the data expected
by the receiver, where 0 ≤ αi ≤ 1 and �
i αi =1. This
can be achieved by determining a sending pattern for each
server, e.g., as in Figure 1, where each server only sends
packets depicted by the solid rectangles. As each packet is
sent by only one of the senders, the total amount of data sent
is the same as in a single path case, i.e., our approach does
not increase the overall workload on the network.
Receiver
Pattern: 0112012
Internet
0 1 1 2 0 1 2
0 1 1 2 0 1 2
0 1 1 2 0 1 2
Figure 1. Multi-path streaming.
Sender 0
Sender 1
Sender 2
In general, we assume that the setting and possible adaptation
of these fractions, as the delivery of data progresses,
is done by the receiver, based on its perceived quality of
data and determination of loss characteristics on the various
paths corresponding to the servers. Thus, the receiver
assembles the data from multiple senders and plays it in an
appropriate order.
There are a number of approaches to accomplishing a
multi-path data delivery. We later describe the specific approach
considered in our work. We first note that such paths
do not have to be completely disjoint, i.e., it is sufficient
for them to have disjoint points of congestion or bottlenecks.
Existence of multiple paths with disjoint bottlenecks
includes the following potential benefits.
• Reduction in correlations between consecutive packet
losses. Although a CM application can tolerate some
missing information, a large number of consecutive
packet losses not only contributes to significant degradation
in CM quality but also diminishes ability to correct
such losses through error correction techniques,
e.g., erasure codes. As shown in [20], sending data
through multiple paths can potentially reduce burst
lengths and correlations between consecutive losses
and thus improve the quality of delivered data.
Proceedings of the Second International Conference on the Quantitative Evaluation of Systems (QEST’05)
0-7695-2427-3/05 $20.00 © 2005 IEEE
• Increased throughput. In delivery of continuous media
one can tradeoff the quality of the data with the
amount of compression achieved, i.e., one can reduce
the amount of bandwidth needed to deliver the data at
the cost of its quality. Sending data through multiple
paths potentially increases the amount of (aggregate)
bandwidth available to the application and hence increases
the quality of delivered data.
• Improved robustness and ability to adjust to variations
in congestion patterns. CM applications are often long
lasting (e.g., delivery of a movie might take on the order
of hours). Hence, it is reasonable to expect that
network conditions will change throughout the delivery
of data to a CM application. Since not all paths,
in general, would experience the same traffic patterns
and congestion, sending data through multiple paths
potentially improves the ability to adapt to changes in
network conditions. Moreover, ability to use multiple
paths can results in improved robustness in the face of
failures.
Given these benefits, the use of multiple paths in distributed
(over best-effort wide-area networks) CM applications,
in general, requires consideration of the following research
problems. (This is not an exhaustive list.)
• Determining bottlenecks, joint points of congestion,
and network characteristics in general. To gain the
benefits of multi-path streaming described above, one
must first determine the paths to be used in delivery
of the data. Since it is reasonable to characterize a
path using its bottleneck link [12], what we need to
be able to do is determine whether a number of paths
share points of congestion, i.e., have joint or disjoint
bottlenecks [21, 29]. In addition, it may be useful (for
some approaches to multi-path streaming) to be able
to estimate current capacities and loss characteristics
of these bottlenecks. Although this has not been critical
in our approach thus far, other approaches to multipath
streaming might require fairly accurate estimation
of various network characteristics (refer to Section 3).
These are non-trivial problems, but they are outside the
scope of this discussion. We note that currently we use
[29] in our system for detecting shared points of congestion.
• Effects of redundancy and error erasure schemes.
Some amount of lost data can be reconstructed in CM
applications through the use of redundant information,
e.g., as in FEC [11] techniques. Hence, in constructing
multi-path streaming techniques one should take
into consideration how the erasure codes interact with
multi-path delivery and the effect of redundant information
on the final quality of the data. In our work, we
try to understand the benefits of multi-path streaming
without focusing on the characteristics of a particular
redundancy scheme.
• Media coding scheme integration. Sophisticated media
coding techniques can be adopted to work closely
with the multi-path streaming scheme (refer to Section
3 for a detailed discussion). In order to increase
the flexibility of multi-path streaming deployment,
we again try to gain understanding on the multipath
streaming benefits without focusing on a particular
media coding scheme.
• Load distribution. As multiple paths are utilized between
the set of senders and the receiver, one can assign
different amounts of streaming workload to the
different paths. The goal here would be to optimize
the viewing quality of the resulting CM data and to
perform the load distribution based on the quality of
the various paths. This problem is explored in [1] and
[18] under different path modeling methods. Example
results are also presented in Section 2.
• Adaptation schemes under changes in network conditions.
When network conditions change, one can improve
the quality of CM data delivery by adapting the
load distribution on the different paths (e.g., by sending
less data on the more congested paths).
• Data placement. Proper placement of data on the
servers is an issue in the context of CM applications
delivering pre-stored data, e.g., a video-on-demand application
(in contrast to a video conferencing application
where data is produced “live”). Inappropriate
data placement can adversely effect servers’ performance.
For instance, this can occur due to load imbalance
problems arising from the fact that only specific
parts of the data are being delivered from a particular
server as well as the fact that specific data required
might change over the course of the application, as the
system adapts to congestion patterns in the network.
This in turn reduces the quality of service experienced
by the CM application (in this case due to server rather
than network performance). We note that these problems
can be more severe when adaptation schemes (as
mentioned above) are used.
• Data dispersion. Given that one cannot necessarily
rely on the network layer to provide multi-path routing,
another consideration is how to accomplish the dispersion
of data over multiple paths existing in the network
between a sender and a receiver of data. This may be
an especially important consideration for applications
where data is generated live, e.g., a video conferencing
Proceedings of the Second International Conference on the Quantitative Evaluation of Systems (QEST’05)
0-7695-2427-3/05 $20.00 © 2005 IEEE
application, where it may be necessary to use a collection
of relay hosts or proxies to “force” paths different
from those provided by the network between a sender
and a receiver. In contrast, this may be less of an issue
for applications where data is pre-recorded and can, for
instance, be dispersed to a set of distributed servers in
advance of actual data streaming.
• Need for protocol/network support. Lastly, some
mechanisms for streaming data over multiple paths
might require support from lower layers, such as the
network layer. Of course, in this case, ease of deployment
is an issue. (We do not require such support.)
From the above list, we can see that there are a number
of difficult problems which need to be addressed when employing
multi-path streaming. In Section 2, we give examples
of results we obtained thus far, which indicate that this
is a worth while approach to achieving high quality streaming
over best-effort networks.
We note, however, that when one adopts multi-path
streaming, its potential costs or detrimental effects should
also be considered. For instance, MP streaming might have
an adverse effect on the resulting delay characteristics observed
at the receiver. As a result, it might also require additional
amounts of receiver buffer space. In addition, the
overheads associated with sending data over multiple paths
and then assembling it into a single stream at the receiver
should also be considered. Moreover, the overheads and
complexity due to measurements needed to achieve better
performance with MP streaming should also be considered.
For instance, in our case, we employ detection of shared
points of congestion [29]. Other approaches to MP streaming
might require even more detailed information about the
network (refer to Section 3) which is likely to result in
a need for more “intrusive” and complex measurements.
Lastly, scalability of such measurement schemes is an issue
as well. However, the evaluation of such costs is outside the
scope of this paper.
Lastly, although all the research issues discussed above
are of importance, in our work thus far we have narrowed
the scope by focusing on: (1) delivery of pre-stored video,
e.g., as in video-on-demand applications (in contrast to delivery
of “live” data); (2) application-level schemes (which
are deployable today over the current Internet) 1 ; (3) accomplishment
of multiple paths to the same receiver by distributing
servers across wide-area networks and streaming
data from multiple senders simultaneously; and (4) streaming
over the network issues only (rather than, e.g., consid-
1 That is, we assume the use of best-effort IP-based networks, where a
specific path is used between any pair of hosts (sender and receiver) on the
network and this path is determined by a network-level routing algorithm;
furthermore, our system does not require specific knowledge of the paths,
only the ability to determine whether two paths share a point of congestion,
e.g., by using [29].
ering server-related problems such as the load balancing issues
mentioned above) 2 .
2 Overview of Results
We now give a brief overview of our results thus far, to
illustrate the potential of multi-path streaming in providing
high quality streaming over best-effort networks.
2.1 Performance Metrics
We begin with a description of performance metrics considered
in our work. Firstly, loss rate, PN is the fraction of
lost packets as seen by the receiver when one uses N ≥ 1
paths for CM streaming. Secondly, lag-1 autocorrelation,
the lag-1 autocorrelation function, R[X(t)X(t + δ)], measures
the degree of dependency of consecutive packet losses
as seen by the receiver, where X(t) is a random variable
indicating whether the packet sent at time t is lost or received
properly (depending on the state of the GM) and 1/δ
is the bandwidth requirement (in units of packets/sec) of the
streaming application 3 . Thirdly, burst length of lost packets,
which is the probability mass function of consecutively lost
packets as seen by the receiver. Note that if the lost packets
burst length is large, it can (a) significantly affect the
viewing quality of the CM object and (b) reduce the effectiveness
of an error correction scheme, if some form of an
erasure code is deployed. Lastly, when considering the use
of erasure codes, we also consider mean information loss
rate (MILR), which accounts for loss of media data (i.e., not
including loss of redundant packets). Although this may be
a useful metric for indicating the resulting visual quality of
CM data, it is often not easy to compute analytically.
2.2 Modeling and Analysis of Benefits
In order to first explore the benefits of multi-path streaming
in a more analytical setting, we model the loss characteristics
of a path 4 using a stationary continuous time
Gilbert model (GM) which can characterize the potential
correlations between consecutive packet losses on a network
path. For a GM, the packet loss process along path
k is described by a two state continuous time Markov chain
{Xk(t)} where Xk(t) ∈{0, 1}. If a packet is transmitted
2That is, for the purposes of this discussion we assume that the data is
fully replicated at all servers and hence any server can deliver any fraction
of the CM data.
3A high positive value of R[X(t)X(t + δ)] implies that a lost packet
is very likely to be followed by another lost packet. A high negative value
of R[(X(t)X(t + δ)] implies that a lost packet is likely to be followed by
a successful packet arrival. If the statistics of the consecutive packet losses
are not correlated, then R[X(t)X(t + δ)] = 0.
4Here we essentially assume that a path is characterized by its bottleneck
link.
at time t when the state of path k is Xk(t) =0, then the
transmitted packet is received correctly by the receiver; the
transmitted packet is considered lost if Xk(t) =1.
Using the GM and the above suggested metrics, an initial
work in [20] gave an analytical characterization of when
a multi-path (MP) approach is beneficial, as compared to a
single path (SP) approach. The results indicated that: (1)
in general, multi-path streaming exhibits better loss characteristics
than single-path streaming, (2) use of an erasure
code may not necessarily improve data loss characteristics
in the case of single-path streaming, while multi-path
streaming (with or without use of an erasure code) can improve
data loss characteristics, and (3) lag1-autocorrelation
of multi-path streaming is usually closer to zero than that
of single path streaming, which we believe should also result
in a higher viewing quality of the received CM data.
It was also indicated that the average error burst length in
multi-path streaming is statistically shorter than that of single
path streaming. As an example, consider Figures 2 and
3 which depict the error burst length distribution obtained
from a prototype experiment 5 using single path and multipath
streaming, respectively. These figures illustrate that
multi-path streaming results in shorter error bursts; here, almost
all bursts in multi-path streaming are one packet long.
2.3 Load Distribution
Proceedings of the Second International Conference on the Quantitative Evaluation of Systems (QEST’05)
0-7695-2427-3/05 $20.00 © 2005 IEEE
Given the above stated benefits, one important question
that remains is how to distribute the load among multiple
paths so as to optimize viewing quality. In [20] Golubchik
et al. study MP streaming benefits using a round-robin (RR)
approach to split the CM traffic among the multiple paths.
In contrast, in [1] Golubchik et al. study the problem of optimally
assigning CM traffic to the multiple paths, according
to the path characteristics, such that a certain specified performance
metric is optimized. This is done under the GM
while considering both appropriate optimization objectives
as well as the resulting streaming performance, all with a
high level goal of improving the perceptual quality of the
streamed media. We note that it is not trivial to pick an appropriate
optimization objective for load distribution. If we
simply use one of the first three metrics given in Section
2.1, undesirable effects might arise. For example, optimizing
the lag-1 autocorrelation may result in higher loss rates
observed at the receiver, when the paths are not homogeneous.
Determining and gaining insight into suitable optimization
objectives is therefore important. Since there is a
fundamental tradeoff between the frequency of losses and
the corresponding loss correlations, it is natural to consider
an optimization objective which encompasses both metrics.
There are a number of ways to include both in a single met-
5 In this prototype, actual data streams are used in experiments; however,
losses are modeled using the GM.
Optimal Load on Path 1 (α (1) )
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
Figure 2. Single path prototype. Figure 3. Dual path prototype.
(1)
Loss Rate on Path 1 (π ) = 0.05
1
0.2
best path
brute force on simulation
optimizing L1AC
0.1
0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
(2)
Loss Rate on Path 2 (π )
1
2
optimizing mean burst length
optimizing MLR × MBL
Figure 4. Optimal Load.
ric. In [1], Mean Loss Rate (MLR) × Mean Burst Length
(MBL) is used as an example of such an optimization objective.
Consider the corresponding example load distribution
results 6 [1] presented in Figure 4, which illustrates an
optimal α ∗ 1 as a function of loss rate on Path 2. (In this
experiment, the Path 1 loss rate is 5% and the redundancy
overhead (due to FEC) is 18.75%.) To compare the effects
of different objective functions on α ∗ when FEC is used,
an optimal traffic load distribution is also computed using
MILR as the optimization objective by a brute force search
on simulation results. (That is, all results other than those
based on MILR are computed analytically here). The corresponding
information loss rates, based on the load distribution
results of Figure 4, are depicted in Figure 5. From
6 Unless otherwise stated, MP streaming is performed on two paths with
SP streaming using the “best path”, i.e., the one with the lower loss rate
(the terms SP and Best Single Path streaming are used interchangeably).
Each experiment is repeated 10 times with different random seeds, and the
results are reported with 95% ± 10% confidence intervals.
Information Loss Rate at Optimal α (1)
5
4.5
4
3.5
3
2.5
2
1.5
1
0.5
(1)
Loss Rate on Path 1 (π ) = 0.05
1
best path
brute force on simulation
optimizing L1AC 2
optimizing mean burst length
optimizing MLR × MBL
Proceedings of the Second International Conference on the Quantitative Evaluation of Systems (QEST’05)
0-7695-2427-3/05 $20.00 © 2005 IEEE
0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
(2)
Loss Rate on Path 2 (π )
1
Figure 5. Information Loss Rate at Optimal.
these figures and results based on other settings [1], the following
observations can be made: (1) When Path 1 loss
rate is relatively low (e.g., 5%), optimal load distribution
on infomation loss increases relatively smoothly as a function
of loss rate on Path 2; it eventually flattens out as it
reaches the SP setting. (2) When Path 1 loss rate is low, optimizing
on MLR×MBL results in a load distribution closer
to the optimal based on MILR. (3) When Path 1 loss rate
is higher, e.g., 10%, the MILR-based optimal is closer to
the load distribution given by SP streaming. (4) If a reasonable
amount of FEC redundancy is used, optimizing on
MLR×MBL gives good performance, i.e., lower information
loss rate, where an MLR×MBL system operates closer
to the MILR-based optimum. We note that when the path
loss rate is relatively high and the corresponding amount of
redundant information is insufficient, using SP streaming
may result in better performance.
Another interesting experiment in [1] compares the ef-
Path2Loss k=8 k=16 k=32 k=64 k=128
0.5% (i) 98.75% 40.63% 18.44% 9.22% 5.00%
0.5% (ii) 77.50% 41.88% 21.88% 13.75% 9.06%
0.5% (iii) 86.25% 38.75% 18.75% 9.38% 5.16%
5% (i) ≥ 1000% 147.50% 58.75% 31.88% 19.92%
5% (ii) 97.50% 55.63% 31.25% 20.31% 14.14%
5% (iii) 97.50% 55.63% 31.25% 20.31% 14.14%
10% (i) ≥ 1000% 147.50% 58.75% 31.88% 19.92%
10% (ii) 128.75% 68.75% 40.63% 26.72% 19.53%
10% (iii) 128.75% 68.75% 40.63% 26.72% 19.53%
Figure 6. Overhead needed to achieve information
loss rate of ≤ 0.1%; Path 1 loss rate = 5%.
fect of redundant information on SP (case i) and MP streaming,
where a simple round-robin approach (case ii) and
the MBL×MLR-based optimization approach (case iii) are
used in the case of MP streaming. Suppose path 1 has a
fixed loss rate of 5%, and three scenarios where Path 2 has
0.5%, 5%, and 10% loss rates are studied. For SP streaming,
the better path (with lower loss rate) is chosen. The
objective of this experiment is to find the minimum FEC
overhead needed to achieve a required quality of service,
which for this experiment we define as a mean information
loss rate of ≤ 0.1%. (Other values have been tried,
and the results are qualitatively similar.) Figure 6 gives the
corresponding mean FEC overhead requirements 7 ; these results
indicate the following. Performance of MLR×MBL is
similar to that of RR when the difference in path loss rates
of the two paths is not large. This is due to the fact that
MLR×MBL results in even traffic splitting in these cases,
which is also shown in Figure 4. Moreover, SP, RR, and
MLR×MBL have different “best” performance ranges. For
example, when path 2 loss rate is 0.5%, SP performs best
(requires the lowest FEC overhead) when the FEC group
size is larger (k ≥ 32). For the same path loss rate,
MLR×MBL performs better than RR when we use k ≥ 16
for FEC. When path 2 loss rate is 5% or 10%, MLR×MBL
(or RR) perform better than SP streaming unless the FEC
group size is quite large.
2.4 A More Expressive Model
The above discussion illustrated that one can obtain improvements
in various metrics by employing MP streaming.
This was shown mainly under the GM; we refer to this
model as a “conventional Gilbert model” in the remainder
of the paper. One major limitation of using a conventional
Gilbert model is that the loss process of a path is independent
of the bandwidth requirements of the streaming appli-
7 This is averaged over 10 experiments; the results are reported with
95% ± 11% confidence intervals. For each scenario, different FEC group
sizes are tested by varying k, the number of packets in a FEC group.
Proceedings of the Second International Conference on the Quantitative Evaluation of Systems (QEST’05)
0-7695-2427-3/05 $20.00 © 2005 IEEE
Measured Loss Rate (%)
40
35
30
25
20
15
10
5
0
120 200 400 600 800 1000 1200
Application’s Sending Rate (pkts/sec)
Figure 7. Data loss rate vs. sending rates.
cation — specifically, the loss rate as viewed by the receiver
is fixed, independently of the sending rate of the application.
To illustrate the potential dependence of the loss rate on
the application’s bandwidth requirements, consider the data
loss rate as a function of an application’s sending rate depicted
in Figure 7; this was obtained through an Internet
experiment8 . This figure illustrates the achieved loss rate
for each experimental setting (i.e., fraction of lost packets as
measured at the receiver), and it supports the hypothesis that
a conventional Gilbert model may not be sufficient for characterizing
the loss process of a path. Given this evidence,
[18] proposes to use a functional Gilbert model (FGM) as a
more general approach to characterizing the bursty loss nature
of a path as well as its dependency on an application’s
bandwidth requirements. Specifically, let λ denote an application’s
average sending rate, in units of packets per second.
For a stationary continuous time FGM, the packet loss process
along path k is described by a two state continuous time
Markov chain {Xk(t)} where Xk(t) ∈{0, 1}. Similarly to
the conventional Gilbert model’s definition, if a packet is
transmitted at time t when the state of path k is Xk(t) =0,
then no packet loss occurs; the transmitted packet is considered
lost if Xk(t) =1. The transition rate from state 0 to
state 1 takes a functional form of F(λ). The transition rate
from state 1 to state 0 also takes a functional form of B(λ).
It is also assumed in [18] that F(λ) and B(λ) are continuous
and furthermore that F(λ) is a non-decreasing function
of λ and B(λ) is a non-increasing function of λ. Intuitively
these assumptions make sense, and hence, in practice, they
should not be restrictive.
8 In this experiment UDP packets (1400 byte) were transmitted from
Hong Kong to the USA, using a number of rates from 120 pkts/sec (around
1.34 Mbps) to 1200 pkts/sec (around 12.8 Mbps), with a step size of 120
pkts/sec interval. For each sending rate, the streaming experiment was
carried out for 6 minutes, while measuring the corresponding achieved loss
rate at the receiver. The experiment was carried out during daytime in
Hong Kong, which corresponds to nighttime on the West Coast of the USA.
Similar experiments were also performed using NS2 [22], and the results
were qualitatively similar.
2.5 Load Distribution Revisited
We now revisit the load distribution problem under the
FGM. Consider the following example result [18] with a
120 pkts/sec bandwidth requirement. We first consider a
system with two heterogeneous paths wherein F1(b) =
0.4 × b, B1(b) = 21000/b, F2(b) = 0.0667 × b, and
B2(b) = 28500/b, as an illustration. (Note that given the
same packet rate on a path, path 2 has a better loss characteristic
than path 1.) Under the best single path streaming
approach (i.e., path 2 in this case), one can achieve a loss
rate of 3.259%. Using 2-path streaming, we reduce the loss
rate to 1.814%, with α ∗ =[0.260, 0.740].
We then add one more path to the system wherein
F3(b) =0.233 × b and B3(b) = 24750/b (i.e., path 3 has
loss characteristics in-between paths 1 and 2). With the addition
of this path we reduce the loss rate further to 0.976%,
with α ∗ =[0.190, 0.541, 0.269]. This example illustrates
that the traffic splitting flexibility of the multi-path approach
provides us the opportunity to reduce the mean loss rate of
a streaming application to a point which would not be possible
with a single best-path type approach. Note that the
additional path which was not present in the 2-path example
is not the best of the three, yet it allows us to reduce the
loss rate further.
The above results were obtained analytically and without
the use of erasure codes. We now consider simulation
results 9 [18] where an error erasure code is used to reconstruct
lost packets. Here the bandwidth requirements of the
streaming application are increased by 12.5% due to the
added overhead of the erasure code. Note that this overhead
is the same for SP and MP streaming. The corresponding
optimal load distributions are obtained based on
the increased packet sending rate (due to the overhead).
The same three heterogeneous paths are considered, i.e.,
paths 1, 2, and 3 described above, where Table 1 illustrates
the corresponding information loss rate for best path, two
paths (with path 1 and 2), and three paths streaming. Load
distribution among the paths is performed by optimizing
packet loss rate as well as in a round-robin manner.
These results and further experiments given in [18] indicate
that adoption of an erasure code, in most cases, can
reduce the information loss rate. However, when the packet
sending rate is high, employing an erasure code may have
an adverse effect of increasing the loss rate (i.e., degrading
the loss characteristics of paths). When an additional
path is available, the workload (including the redundancy
overhead) can be spread among paths; this results in better
information loss rate. It is also observed that loss-based op-
9 In these experiments, each data packet has a size of 1400 bytes. For
each path, packet losses are emulated according to the FGM. Each simulation
is analogous to a 24 hour media stream. The results are obtained using
CSIM[23].
timal load distribution among the multiple paths can results
in significantly better system performance than single best
path streaming or the round-robin approach.
2.6 Experiments with Real Data
Proceedings of the Second International Conference on the Quantitative Evaluation of Systems (QEST’05)
0-7695-2427-3/05 $20.00 © 2005 IEEE
Chow et al. [18] discuss an experiment with real data using
a multi-path multimedia streaming system prototype using
the performance metrics given in Section 2.1 as well as
the resulting visual quality under different load assignment
methods. (A detailed description of the prototype can be
found in [2].) Although packet losses are still emulated using
the FGM, the MPEG stream and their processing is real,
hence we are able to depict the resulting visual quality in addition
to computing performance metrics 10 . Two senders,
whose paths to the receiver have different loss characteristics,
are used in this experiment. Path 1, which has better
loss characteristics, is described by: F1(b) =0.0667 × b,
B1(b) = 28500/b, and path 2, which has worse loss characteristics,
is described by: F2(b) = 0.4 × b, B2(b) =
21000/b, where b = 192 packets/sec. In this case, when
one optimizes on the loss rate, α ∗ =[0.741, 0.259] is obtained.
Four cases with different traffic splitting vectors are studied:
(Case 1) single path with better loss characteristics
(using path 1), (Case 2) single path with worse loss characteristics
(using path 2), (Case 3) dual path using RR traffic
splitting, and (Case 4) dual path using optimal traffic
splitting vector α ∗ . Packet loss statistics are measured at
the receiver throughout the entire streaming process. Video
frames are transcoded to JPEG files to allow a visual quality
inspection.
Test Loss Rate Avg. Burst Length Info.
Case before FEC before FEC Loss Rate
1. 6.569% 1.762 1.688%
2. 35.630% 2.439 35.642%
3. 6.955% 1.034 1.000%
4. 3.403% 1.253 0.090%
Table 2. Prototype experiments.
Table 2 shows the average statistics measured for each
test case. For the statistics before FEC, the statistics of
the resultant packet receiving sequence after merging packets
from different paths are measured. Figures 8(a), 8(b),
and 8(c) depict the quality of the video frame sequence
along playback time for Cases 1, 3, and 4, respectively 11 .
For each case, the transcoded video frame sequence is examined,
and a vertical spike is depicted in the figure where
10 A 150 seconds MPEG1 file which requires a playback rate of 174.5
Kbps (approximately 170 packets/sec with a packet size of 1024 bytes) is
streamed. A FEC corresponding to an overhead of 12.5% is used, which
increases the packet sending rate to around 192 packets/sec.
11 All video frames in Case 2 are damaged, and thus the results for that
case are not presented.
Rate 2-path 3-path 2-path Opt. Info. 3-path Opt. Info. Best SP 2-path RR 3-path RR
(pkts/sec) α ∗ α ∗ Loss Rate Loss Rate Info. Loss Rate Info. Loss Rate Info. Loss Rate
60 [0.260,0.740] [0.191,0.540,0.269] 0 0 0 0 0
120 [0.260,0.740] [0.190,0.541,0.269] 0.001% 0 0.177% 0.057% 0
360 [0.260,0.740] [0.190,0.541,0.269] 15.055% 5.301% 26.754% 26.128% 12.256%
the video frame at that playback time is damaged (i.e., the
picture exhibits a “blocking effect”). From these results, it
(a) Case 1: Single Best Path Streaming.
Frame sequence with 14.591% damaged frames.
begin end
time
(b) Case 3: Streaming under Dual path with RR Splitting.
Frame sequence with 9.517% damaged frames.
begin end
time
(c) Case 4: Streaming under Dual Path with Optimal Splitting.
Frame sequence with 1.278% damaged frames.
begin end
time
Figure 8. Video quality along playback time.
is observed that:
1. Loss rates (after FEC) significantly affect visual quality:
When one relates the after FEC loss rate in Table
2 with the frame sequences quality indication in
Figure 8, it is found that higher loss rate after the
FEC operation, i.e., information loss rate, corresponds
to poorer video quality, i.e., higher percentage of the
video sequence is damaged. When the information
loss rate is extremely high, for example 35.642% in
Case 2, all the frames in the video sequence are damaged,
i.e., the video is basically ruined and unviewable.
2. Introducing FEC may cause adverse effects: Improper
adding of FEC may not improve the resulting information
loss rate. In Case 2, it is found that loss rate
before the FEC operation is less than the loss rate after
the FEC operation. Adding FEC increases the loading
along a congested path which may degrade the resulting
loss characteristics and consequently increase data
packet losses.
3. FEC performs better under multi-path streaming: As
suggested in [20], multiple paths can reduce the lag-
1 auto-correlation and shorten the error burst length;
this enhances the error correction capability of FEC.
The loss rates before FEC for Case 1 and Case 3 are
6.569% and 6.955%, respectivly. Using the best path
Table 1. Optimization of load distribution revisited.
in Case 1 can give a lower loss rate before the FEC
process. However, after the error correction process,
information loss is 1.688% for Case 1 and 1.000% for
Case 3. The visual quality, which is strongly related to
the information loss rate, is better in Case 3. It shows
that simply using the best single path for video streaming
may not be a good approach.
4. A path with worse loss characteristics can be used to
improve the overall performance: Although path 2 is a
worse path than path 1, it can still be used to share a
fraction of the workload to improve the overall quality
of the received video. From the statistics in Table 2 and
the corresponding video output, the two multi-path test
cases give better performance than the two single path
test cases.
5. Using optimal traffic can result in better performance:
In the experiments, when one assigns traffic loads according
to an optimized traffic splitting vector, i.e.,
Case 4, the loss rate is the lowest both, before and after
FEC. It shows that simply splitting the traffic evenly
between paths may not result in the best use of available
multiple paths (although it may lower the packet
loss correlations).
3 Related Work
Proceedings of the Second International Conference on the Quantitative Evaluation of Systems (QEST’05)
0-7695-2427-3/05 $20.00 © 2005 IEEE
Multi-path streaming and exploitation of path diversity
has attracted much attention recently, and [7] provides a
broad overview of the general area. Here we give a brief
survey of existing work on this topic focusing on those
works which either consider loss characteristics or can be
deployed over best-effort networks, as these are considerations
in our work as well. As mentioned earlier, [20] illustrates
the potential benefits of using multi-path streaming
to improve the quality of the delivered CM data as compared
to single path streaming. This is done by illustrating
lower loss burst lengths and lower correlations in consecutive
packet losses. A later work [1] studies the load distribution
problem in multi-path streaming and shows that
both the packet loss rate and the loss correlations are important
when choosing an optimization objective. However,
all these works are performed under the assumption that the
application’s sending rate does not affect the loss rate on a
path (i.e., the conventional Gilbert Model is used). In con-
trast, [18] illustrates the utility of considering an application’s
sending rate and the resulting effects on the loss characteristics
of the CM streaming application.
Other works try to achieve multi-path streaming with assistance
of the lower layers, e.g., [19] focuses a protocol
which utilizes bandwidth available on multiple paths and redistributes
workload according to congestion detected by a
receiver. However, it requires network layer knowledge as
well as a centralized routing server to compute the multipath
routing which minimizes the number of shared links
while satisfying the bandwidth requirement. Given information
about the underlying network graph, [17] proposes
multi-path routing heuristics for unicast and multicast scenarios
to achieve high bandwidth and low delay for video
traffic. A data scheduling algorithm at the server is also
proposed to minimize the end-to-end delay. Similarly, given
network link information, [10] discusses a heuristic solution
to find a set of paths that minimize the streaming distortion
for a Multiple Description (MD) coded video stream. These
works assume significant knowledge (i.e., link bandwidth
and delay) and support (i.e., ability to control routing paths)
from the underlying network. In contrast, our approach only
deals with the end-hosts of the network and hence allows
easy deployment on the Internet. Our only requirement is
that chosen paths do not share points of congestion, which
can be detected at the end-hosts using schemes such as [29].
Also, we focus on packet loss characteristics (e.g., loss correlations)
rather than bandwidth and delay.
Sophisticated media coding techniques (e.g., MD coding)
have been developed for use with multi-path streaming
[4, 8, 6, 9, 5, 24, 30, 13, 14, 16, 15]. The basic idea
is to partition the media into multiple roughly equally important
independently decodeable bitstreams (descriptions).
Each of them contains complementary information and is
sent through different paths, such that media quality can be
improved as the number of received descriptions increases.
In this context, issues such as dealing with heterogeneous
path bandwidth constraints [8], rate-distortion optimization
[24, 13, 14], coding efficiency, adaptation schemes [30, 16],
etc. are explored. Specifically, in [6, 9], a model based on
the conventional GM is also used in the analysis and path
selection for MD coded multi-path streaming. In our work,
we focus on a more general study of multi-path streaming
without the use of a specific coding technique - an advantage
being that it can then be made to work well with many
coding techniques or media codecs. Also, our work adopts
a more expressive model for characterizing the loss characteristics
of paths between senders and a receiver. Developing
an adaptation scheme which reacts to time-varying path
characteristics is our ongoing effort.
The work in [27] proposes the use of multi-servers for
video streaming on the Internet. Their later work [26] extends
the scheme by considering the use of FEC. In their
work, they focus on designing a receiver-driven transport
protocol which includes (a) a rate allocation algorithm (i.e.,
how receiver split the traffic load onto multiple paths in
order to minimize the probability of media packet loss by
taking network bandwidth, channel characteristics and FEC
parameters into account) and (b) a packet partitioning algorithm
(which ensures non-overlapping packet sending and
minimizes startup delay). In [25], they examine the case
where the last mile connection is the bottleneck and employ
multi-path streaming otherwise. In these proposals, conventional
GM is adopted while a more expressive model is
used in our work. Besides focusing on the loss rate, we also
propose an optimization approach using other loss characteristics,
e.g., lag-1 autocorrelation. Their later work [28]
extends the idea of using FEC with path diversity on an
overlay framework. The motivation is to emulate multiple
sources by the use of an overlay network, which is similar
to [3] but using multiple redundant paths simultaneously.
They propose a heuristic scheme to select redundant paths.
It is suggested that 10% of the network nodes participating
in the overlay is enough for providing sufficient redundant
paths. Our approach can also be implemented on an overlay
structure as well as be made to achieving multiple paths by
using relay nodes. In fact, we have a P2P prototype implementation
of our approach [2].
Best-path type approaches have also been studied. For
instance, [31, 32] perform path switching to select the best
path by estimating the “goodness” of a path from the perspective
of a video and a VoIP stream. However, they do not
exploit the benefits of path diversity (e.g., reduced correlation
in packet losses, workload distribution, etc.) and are
similar to the best single path case discussed in our work.
4 Conclusions
Proceedings of the Second International Conference on the Quantitative Evaluation of Systems (QEST’05)
0-7695-2427-3/05 $20.00 © 2005 IEEE
In this paper we focused on multi-path streaming as an
approach to providing high quality streaming over the Internet.
We discussed evidence that such an approach has
promise. Although much research still remains on this
topic, we believe that this evidence indicates that there is
hope for being able to stream continuous media over the
Internet via the multi-path approach, with high quality and
without huge costs.
To this end (and to verify our analytical and simulation
results under real world conditions), we designed and built a
P2P-based multi-path streaming prototype [2], currently being
used for streaming video experiments between UMD,
CUHK, and USC. We chose to use a P2P network as the
supporting architecture for our prototype because of the parallelism
between simultaneous downloads in P2P systems
and multi-path streaming. Currently, we are working on Internet
based experiments using this prototype.
References
[1] B. Abdouni, W. C. Cheng, A. L. Chow, L. Golubchik, W.-J.
Lee, and J. C. Lui. Multi-path streaming: Optimization and
performance evaluation. In SPIE Conference on Multimedia
Computing and Networking (MMCN), volume 5680, pages
216–227, San Jose, California, January 2005.
[2] B. Abdouni, W. C. Cheng, A. L. Chow, L. Golubchik, and
J. C. Lui. Picture-perfect streaming over the internet: Is
there hope? IEEE Communication Magazine special issue
on Proxy Support for Streaming on the Internet, 42(8):72–
79, Augest 2004.
[3] D. G. Andersen, H. Balakrishnan, M. F. Kaashoek, and
R. Morris. The case for resilient overlay networks. In HotOS
VIII, Schloss Elmau, Germany, May 2001.
[4] J. Apostolopoulos. Reliable video communication over
lossy packet networks using multiple state encoding and
path diversity. In Visual Communications and Image Processing,
January 24-26, 2001.
[5] J. Apostolopoulos, W. Tan, and S. Wee. Performance of a
multiple description streaming media content delivery network.
In IEEE International Conference on Image Processing
(ICIP), September 2002.
[6] J. Apostolopoulos, W. Tan, S. Wee, and G. Wornell. Modeling
path diversity for multiple description video communication.
In IEEE Inter. Conf. on Acoustics, Speech, and
Signal Processing (ICASSP), May 2002.
[7] J. Apostolopoulos and M. Trott. Path diversity for enhanced
media streaming. IEEE Communication Magazine
special issue on Proxy Support for Streaming on the Internet,
42(8):80–87, Augest 2004.
[8] J. Apostolopoulos and S. Wee. Unbalanced multiple description
video communication using path diversity. In IEEE International
Conference on Image Processing (ICIP), October
2001.
[9] J. Apostolopoulos, T. Wong, W. Tan, and S. Wee. On multiple
description streaming with content delivery networks. In
IEEE INFOCOM, June 2002.
[10] A. Begen, Y. Altunbasak, and O.Ergun. Fast heuristics fro
multi-path selection for multiple description encoded video
streaming. In IEEE Conference on Multimedia and Expo,
pages 517–520, Baltimore, MD, July 2003.
[11] R. E. Blahut. Theory and Practice of Error Control Codes.
Addison Wesley, January 1983.
[12] J.-C. Bolot, S. Fosse-Parisis, and D. Towsley. Adaptive
FEC-Based Error Control for Internet Telephony. In INFO-
COM, 1999.
[13] J. Chakareski and B. Girod. Rate-distortion optimized
packet scheduling and routing for media streaming with path
diversity. In IEEE Data Compression Conference, Snowbird,
UT, April 2003.
[14] J. Chakareski and B. Girod. Server diversity in ratedistortion
optimized media streaming. In IEEE International
Conference on Image Processing, Barcelona, Spain,
September 2003.
[15] J. Chakareski, S. Han, and B. Girod. Layered coding
vs. multiple descriptions for video streaming over multiple
paths. In ACM Multimedia, Berkeley, CA, November 2003.
Proceedings of the Second International Conference on the Quantitative Evaluation of Systems (QEST’05)
0-7695-2427-3/05 $20.00 © 2005 IEEE
[16] J. Chakareski, E. Setton, and B. Girod. Video streaming with
diversity. In IEEE Conference on Multimedia and Expo,
pages 9–12, Baltimore, MD, July 2003.
[17] J.-C. Chen, S.-H. Chan, and V. Li. Multipath routing for
video delivery over bandwidth-limited networks. IEEE
Journal on Selected Areas in Communications special issue
on Design, Implementation and Analysis of Communication
Protocols, 22(10):1920–1932, December 2004.
[18] A. L. Chow, L. Golubchik, J. C. Lui, and A. W.-J. Lee.
Multi-path Streaming: Optimization of Load Distribution.
Accepted for publication in Journal of Performance Evaluation.,
2005.
[19] H. Chu and K. Nahrstedt. Dynamic multi-path communication
for video traffic. In Hawiian International Conference
on System Science, Hawaii, January 1997.
[20] L. Golubchik, J. C. Lui, T. F. Tung, A. L. Chow, W.-J. Lee,
G. Franceschinis, and C. Anglano. Multi-path Continuous
Media Streaming: What are the Benefits? Performance
Evaluation, 49:429–449, September 2002.
[21] K. Harfoush, A. Bestravos, and J. Byers. Robust identification
of shared losses using end-to-end unicast probes. In
The 6th IEEE International Conference on Network Protocols
(ICNP), Osaka, Japan, October, 2000.
[22] http://www.isi.edu/nsnam/ns/. The Network Simulator - ns-
2.
[23] http://www.mesquite.com/. CSIM18.
[24] Y. J. Liang, E. Setton, and B. Girod. Channel-adaptive video
streaming using packet path diversity and rate-distortion optimized
reference picture selection. In IEEE Fifth Workshop
on Multimedia Signal Processing, St. Thomas, Virgin Islands,
December 2002.
[25] T. Nguyen, P. Mehra, and A. Zakhor. Path diversity and
bandwidth allocation for multimedia streaming. In ICME,
Baltimore, MD, USA, July 6-9, 2003.
[26] T. Nguyen and A. Zakhor. Multiple sender distributed video
streaming. IEEE Tran Multimedia, pages 315–326, April
2004.
[27] T. Nguyen and A. Zakhor. Distributed video streaming over
internet. In SPIE Conference on Multimedia Computing and
Networking, San Jose, California, January 2002.
[28] T. Nguyen and A. Zakhor. Path diversity with forward error
correction (pdf) system for delay sensitive applications over
the internet. In IEEE InfoCom, San Francisco, California,
March 2003.
[29] D. Rubenstein, J. Kurose, and D. Towsley. Detecting
shared congestion of flows via end-to-end measurement.
IEEE/ACM Transactions on Networking, 10(3):381 – 395,
June 2002.
[30] E. Setton, Y. Liang, and B. Girod. Adaptive multiple description
video streaming over multiple channels with active
probing. In IEEE Conference on Multimedia and Expo,Baltimore,
MD, July 2003.
[31] S. Tao and R. Guerin. Application-specific path switching:
A case study for streaming video. In ACM Multimedia,New
York, October 2004.
[32] S. Tao, K. Xu, A. Estepa, T. Fei, L. Gao, R. Guerin,
J. Kurose, D. Towsley, and Z.-L. Zhang. Improving voip
quality through path switching. In IEEE InfoCom, Miami,
March 2005.
