LOSS CONTROL THROUGH THE COMBINATION OF BUFFER

MANAGEMENT AND PACKET SCHEDULING

Yan Bai, Mabo Robert Ito

Department of Electrical and Computer Engineering , University of British Columbia

2356 Main Mall Vancouver, BC V6T 1Z4 Canada

Keywords: Quality of Services, Packet Scheduling, Buffer Management

Abstract: Conventional Quality of Service (QoS) control techniques are designed for achieving network-level QoS

objectives. Due to the large differences between network-level and application-level QoS properties, these

techniques cannot provide desirable QoS for video users. Previous work has been conducted to design a

packet scheduling approach where application requirements and network-level QoS objectives are addressed

simultaneously. In this paper, the packet scheduling approach is integrated with a buffer management

technique for increasing the numbers of video users with QoS satisfaction. The effectiveness of the

proposed technique is demonstrated through simulations.

1 INTRODUCTION

Quality of Service (QoS) can be achieved through

the use of packet scheduling techniques. Majority of

the packet scheduling methods has been designed for

allocation of a minimum bandwidth to each flow

that crosses a link and provision of throughput and a

delay bound [Zhang, 1995]. For example, Weighted

Fair Queuing (WFQ) [Parekh and Gallager, 1993]

assigns a weight to each flow. The weight logically

specifies how many bits to transmit each time the

router services that particular flow; this controls the

percentage of the link's capacity for that flow. The

weight depends on the request rate of each flow and

how many other flows are sharing the link. Each

request rate is assured if the sum of all the request

rates is smaller than the link capacity; otherwise, the

allocated rate is the ratio of the request rate of the

flow to the sum of the request rate of the backlogged

flow at that instant. Another representative

scheduling mechanism is Class-based Queuing

(CBQ) [Floyd and Jacobson, 1995]. CBQ is a

hierarchical link-sharing mechanism. It partitions

network bandwidth among the different traffic

classes, in which a higher percentage of bandwidth

is allocated to the important traffic class.

These scheduling mechanisms are very efficient

from a network perspective. However, they are

inadequate from the viewpoints of application users,

such as video users. The reason is that they are not

designed to provide application-level QoS but QoS

from the standpoint of a network. However, for

video applications, there are performance gap

between application-level QoS and network-level

QoS. Specifically, there is no linear relationship

between visual quality and bit-rate. Some bits may

be more important than others. In other words,

perceived video quality will generally be dependent

on the data rate and the data content. Current packet

scheduling techniques do not use content

information in the data; rather, they treat all the data

in the same manner. Another common problem with

existing scheduling techniques is that they consider

dropping data that misses the deadline and do not

consider data loss due to buffer overflow [Dovrolis

and et.al.]. To overcome these limitations, we

proposed a new packet scheduling scheme, called

MPAPS [Bai and Ito, 2003]. The idea of MPAPS is

to make enough available buffer space before

incoming packets arrive by appropriately scheduling

queued packets to exit the router. Thus, incoming

packets will be admitted into the router and the

packet loss due to buffer overflow will be reduced.

Also, MPAPS considers the characteristics of video

data and users’ requirements, as well as network

QoS parameters in the control of service time and

250

Bai Y. and Robert Ito M. (2004).

LOSS CONTROL THROUGH THE COMBINATION OF BUFFER MANAGEMENT AND PACKET SCHEDULING.

In Proceedings of the First International Conference on E-Business and Telecommunication Networks, pages 250-255

DOI: 10.5220/0001401002500255

 SciTePress

service order of the queued packets. Consequently,

the perceived QoS of video users is improved.

In MPAPS, when the input buffer at a router is

full, arriving packets are dropped. Thus, it is likely

to introduce arbitrary loss distribution between

videos and different parts of a video. As a result, the

numbers of video users with QoS satisfaction is

reduced because the locations at which the losses

occur can have a significant effect on the QoS

perceived by the user. For example, a loss of several

consecutive packets in a frame may be imperceptible

to the user whereas the same loss rate distributed

over different video frames can largely degrade

visual quality [Ito and Bai, 2002]. Therefore, an

Enhanced MPAPS (E-MPAPS) is needed. In the E-

MPAPS, a new buffer management at the input

buffer is integrated to the original MPAPS in order

to control loss distribution at the input buffer.

This paper investigates the performance of E-

MPAPS scheme. Section 2 describes the scheme.

Section 3 presents the simulation results, and

Section 4 concludes the paper.

2 SCHEME DESCRIPTION

E-MPAPS targets MPEG video due to the

abundance of existing videos of MPEG format. It is

composed of packet scheduling and buffer

management (Figure 1).

The buffer management at the input buffer of a

router is designed to maintain high number of videos

with QoS satisfaction. Therefore, it accomplishes the

following two criteria:

1) Selection of which video to be rejected when

congestion occurs such that every video achieves a

loss performance at a level commensurate with its

individual expectations. It other words, no

overservicing or underservicing of a particular video

stream occurs. Thus, network utilization is

maximized.

2) Determination of how much data from the

selected streams should be discarded during the

periods of congestion in order to maximize the video

quality in the presence of packet loss.

Sender Receiver

Sender

Figure 1: The E-MPAPS Scheme

The pseudo code of buffer management is listed

in Figure 2.

LOW: the buffer length threshold at

which B-packets start being dropped.

HIGH: the buffer length threshold at

which P-packets start being dropped.

Len: the buffer length.

Size: the buffer size.

E-frame: a partially discarded frame.

PLRi: the acceptable packet loss ratio

of video stream i.

Wi: weight parameter for video stream

i, it is set inversely proportional to

its packet loss tolerance.

∆: the difference between initial and

update values of LOW.

if (packet == E-frame)

drop();

if (Len == Size)

drop();

else if (Len > HIGH ){

if((packet == first P-packet )

|| (packet == B-packet)){

drop();

}

else

accept();

}else if (Len > LOW){

Input Virtual MPAPS

Buffer Buffer Scheduler

LOSS CONTROL THROUGH THE COMBINATION OF BUFFER MANAGEMENT AND PACKET SCHEDULING

251

if ((packet == B-packet) ||

(Len > Wi*Size ))

drop();

else

accept();

}else accept();

/* In the drop () procedure, LOW

increases by one when an I- or P-packet

gets accepted subject to ∆ remaining

greater than zero. */

Figure 2:.Pseudo Code of Buffer Management at Router

Input Buffer

Packet Scheduling is based on the original

MPAPS scheme. It is designed for per-flow queue

router structure. The E-MPAPS scheme thus

constructs virtual buffers in order to use MPAPS

with the above buffer management method. Each

virtual buffer holds a video stream. The virtual

buffer size (Bv) is computed according to the

following equation.

= W

*Size (1)

Due to the use of virtual buffer E-MPAPS has

one distinguished feature: it does not provide strict

isolation between different videos and video sources

could use the buffer reserved to others when load

level at the router is low.

MPAPS first maps the videos into the groups

with different transmission priorities based on their

upcoming packet type and current loss performance

(SL), and then a specific transmission schedule for a

video stream in a selected group is adaptively set to

respond instantaneously to needs based on the

Adaptive Priority Index (API). Here, the SL is

defined as ratio of the actual packet loss and the

maximum allowable packet loss, and the API is

defined as the product of SL and the normalized

length of an virtual input buffer holding stream i.

The rationale behind MPAPS is that the drop

probability of I packets will be lower than that of P

packets, which in turn, will be less than that of B

packets. Moreover, the videos that have a worse loss

performance than expected receive expedited and

more servicing, whereas videos that have

satisfactory or even better loss performance receive

slow and less servicing. Therefore, all the videos

will be transmitted with more acceptable loss

targets. Further details can be referred in [Bai and

Ito,2003].

3 RESULTS AND DISCUSSION

To demonstrate the advantages of E-MPAPS, we

compare its performance, including packet loss rate

and I-frame error rate with that of the original

MPAPS, and first-come-first-serve scheduling

(FCFS) schemes. FCFS is widely used in the current

Internet routers due to its simple implementation. In

FCFS, the incoming packets are accepted in order of

arrivals. The simulation details are presented in the

following.

Sources: real MPEG-1 video traces where the

number of bits per frame used by the MPEG coder is

described [http://www3.informatik.uni-

wuerzburg.de/MPEG/].

Parameters used in the simulation:

Fixed:

• Packet Size: 1500 bytes or less

• Simulation Time: 40 minutes

• Video Starting Interval: 60 seconds

• Output Link: 100 Mbps

• Size: 150KB

• LOW: 0.90 [a threshold value of 0.90

means that the buffer length threshold is

90% of the buffer size (in packets)]

• HIGH: 0.95

• PLRi: 3% for first half of videos and

6% for the others

2) Variable:

• The number of background sources

varies in order to change the load level

at the router.

• The starting sequence of a video stream

was randomly selected in each run. The

results presented in this section show

the final values of the average of

different runs.

• The test-scenarios featured varying

degrees of load level and various traffic

patterns. The presented test results here

are an illustrative comparison of the

differences of the three schemes.

Table 1 and Figures 3 and 4 show the results

obtained for different number of transmitted videos,

namely, 10, 20 and 30 videos.

We see in Table 1 that E-MPAPS produces the

lowest number of videos whose actual packet loss

rate is greater than their maximum allowable packet

loss rate for all the cases.

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

252

Looking at the standard deviation of packet loss

difference for each of the three cases, we find that

the smallest value appears in the E-MPAPS scheme.

It indicates that the loss differences fluctuate slightly

between videos when E-MPAPS is applied. In other

words, for many videos, the actual loss of each

stream is controlled to nearly match the specified

allowable values. Conversely, relative high and too

much loss variation from expected packet loss rate

appears in MPAPS and FCFS. This gives the reason

why E-MPAPS decreases the number of videos with

packet loss beyond the expected values.

The above results can be explained as follows.

FCFS does not have any loss distribution control

mechanism. While MPAPS adaptively adjusts the

probability of transmissions that individual streams

receive: those stream queues that are in danger of a

violation of loss requirements receive servicing

sooner and more frequently. Conversely, those

stream queues that have a lower loss than expected

receive servicing that is delayed and less frequent.

When streams are delivered to a network node

quickly enough to make MPAPS impossible to

adjust the servicing sequence and frequency timely,

unexpected packet loss would occur. It results in an

inequitable loss distribution between videos.

E-MPAPS, except the adaptation mechanism

done by MPAPS, the buffer management allocates

buffer occupancy between videos in a fair manner:

buffer occupancy is inversely proportional to their

loss constraints. Therefore, the losses distribution

among the streams is further enforced, just matching

their individual loss tolerance.

Table 1: Comparison of Packet Loss

Number of videos 10 20 30

E-MPAPS 4 6 6

MPAPS 6 9 12

Number of violating

stream

FCFS 7 11 18

E-MPAPS 0.00091 0.00085 0.00170

MPAPS 0.00102 0.00245 0.00241

Standard Deviation

of ∆

FCFS 0.00614 0.00787 0.00771

Violating stream: actual packet loss rate is greater than their maximum allowable packet

loss rate.

∆: the difference between achieved packet loss and maximum allowable packet loss for each

individual video.

Figures 3 to 4 plot I-frame error rate for the three

schemes when the number of transmitted video (N)

is 10, 20 and 30, respectively. I-frame is defined as

an error frame if one packet in the frame is lost.

From the figures, we obtain the following. For all

the three cases, the I-frame error rate in E-MPAPS is

no more than 0.3%, while around 2-3% in MPAPS.

In FCFS, I-frame error rate is largely increased,

approximately 20% with N=30.

This can be explained as follows. E-MPAPS

includes buffer management mechanism. It detects a

congestion condition by observing when the router’s

input buffer occupancy is close to crossing, or has

crossed, a specified threshold and determines how

best to allocate the available buffer to reduce the I-

packet loss. Instead, MPAPS discards packets

arbitrarily, therefore, most likely distributing packet

loss over all frame types during the congestion

episode. In particular, a lost packet could belong to

an I-frame.

LOSS CONTROL THROUGH THE COMBINATION OF BUFFER MANAGEMENT AND PACKET SCHEDULING

253

Figure 3: I-frame Error Rate for the First Half of Videos

Figure 4: I-frame Error Rate for the Second Half of Videos

Adding buffer management in E-MPAPS

largely decreases I-frame error rate, giving a great

increase in the visual quality of video. It suggests

that the number of video users with QoS

satisfaction is increased. The reason is that the I-

frame error has significant adverse effects on the

perceived video quality. Each video frame is very

large in size, and is thus segmented into a sequence

of IP packets when delivered through an IP

network. The loss of a packet may cause the errors

in a video frame. Furthermore, MPEG video

possesses a frame independent nature. The I-frame

is coded independently. The P-frame and B-frame

are coded by using the closest past I- or P-frame,

and the closest past and future I- or P-frames,

respectively. Therefore, the loss of an I-packet

distorts the whole GOP, which is equivalent to

affecting half a second of video.

4 CONCLUSION

In this paper, a modification of the MPAPS

scheme (E-MPAPS) is proposed. E-MPAPS

introduces an application-aware buffer

management in MPAPS at the input buffer of a

router. It improves the loss distribution between

different parts of a video stream and the

contending videos. Simulation results have shown

0.00%

5.00%

10.00%

15.00%

20.00%

25.00%

N=10

0.10% 1.90% 13.30%

N=20

0.30% 2.62% 19.90%

N=30

0.31% 2.88% 14.10%

E-MPAPS MPAPS FCFS

0.00%

5.00%

10.00%

15.00%

20.00%

N=10

0.10% 1.39% 14.40%

N=20

0.20% 1.60% 16.97%

N=30

0.33% 2.30% 13.12%

E-MPAPS MPAPS FCFS

ICETE 2004 - WIRELESS COMMUNICATION SYSTEMS AND NETWORKS

254

that E-MPAPS significantly improves the number

of videos with high visual quality over MPAPS

and the conventional packet scheduling method,

without the use of data content. E-MPAPS,

however, has a relatively high computational

complexity due to the use of application

requirements in the buffer management.

REFERENCES

Bai, Y. and Ito, M.R., 2003. Packet Scheduling to

Support Loss Guarantee for Video Traffic, In

ICT’03, 10th International Conference on

Telecommunications.

Dovrolis, C., Stiliadis, D. and Ramanathan, P., 2002,

Proportional Differentiated Services: Delay

Differentiation and Packet Scheduling, In

SIGCOMM’02.

Floyd, S., and Jacobson, V., 1995, Link-sharing and

Resource Management Models for Packet Networks,

IEEE/ACM Transactions on Networking, Vol. 3 No.

4, pp. 365-386, August 1995.

Ito, M.R and Bai, Y., 2002, A Packet Discard Scheme

For Loss Control in IP Networks With MPEG Video

Traffic, In ICCS’02, 8th IEEE International

Conference on Communication Systems.

Parekh, A. K. and Gallager, R. G., 1993, A Generalized

Processor Sharing Approach to Flow Control in

Integrated Services Networks: The Single-Node

Case, IEEE/ACM Transactions on Networking, Vol.

1, No. 3, 1993, pp. 344–57.

Zhang, H., 1995 Service Disciplines for Guaranteed

Performance Service in Packet-Switching Networks,

In Proceedings of the IEEE, Vol.83, No.10, pp.1374-

1396, October 1995.

LOSS CONTROL THROUGH THE COMBINATION OF BUFFER MANAGEMENT AND PACKET SCHEDULING

255