Korea Uni Presentation

1

Achieving QoS Efficiently in the Internet in the Presence of

Bursty Self-Similar TrafficBy

Chandana Watagodakumbura PhD

2

Presentation Outline

• Transfer of real-time traffic over the Internet – Significance – Challenges faced

• Intuitive understanding of the problem

• Possible solution directions– Exploiting statistical multiplexing gains

• Understanding statistical multiplexing gains

• Working Through a Solution: Priority Queue with Lower Real-Time Utilisation (PQ-LRTU) framework– Statistical Guarantees for Real-Time Traffic in the Internet

3

Real-Time Traffic Over the Internet - Challenges

• What is the significance of transmitting real-time traffic over the Internet?

– More economical approach to real-time, voice/video data transmission

• What are the challenges involved compared to the use of the Internet for traditional best-effort data transmission?

– Self-Similar or bursty nature of Internet traffic causes performance degradation

– Providing quality of service (QoS) guarantees to end-users

– The need to use resources efficiently thereby minimizing cost

4

Real-Time Traffic Over the Internet - Challenges Contd.

• How can we get an intuitive idea of the challenges involved?

• What does it really mean by bursty traffic?

• How do you measure burstiness?

– Hurst parameter (H)

• What causes internet traffic busty?

• How do you quantify relatively the adverse performance effects of bursty traffic?

– Figure 1 shows (next slide) adverse effects of self-similarity, or burstiness, on average delay of a data flow

– The results were obtained simulating aggregation of a number of traffic flows of different levels of self-similarity (or different Hurst parameter, H, values)

5

Real-Time Traffic Over the Internet - Challenges Contd.

Figure1: Relative variation of mean queue delay with Hurst parameter H and number of micro flows

6

Possible Solution Directions:Statistical Multiplexing Gains

• Why does delay decrease when the number of data flows aggregated are increased?

– Allows statistical multiplexing gains

• To represent possible statistical multiplexing gains in an analytical model, we use binomial distribution

• The mean queue delay is dependant on the number of sources that is active (represented by random variable X) out of a maximum of Nmax sources, at a given point in time

7

Possible Solutions Directions:Statistical Multiplexing Gains Contd.

Table 1:Probability of At Least More Than The Specified Number of Sources Are Active Using Binomial Distribution

Probability

Number of Maximum Sources (Nmax)

1 2 4 8 12

P[X >= Nmax/2] 0.5 0.75 0.6875 0.6367 0.6128

P[X >= 4Nmax/5] 0.5 0.25 0.0625 0.035 0.01930

0.2

0.4

0.6

0.8

1

1.2

0 2 4 6 8 10 12 14

number of micro f low s

prob

abilit

y

50% or more active 80% or more active66.7% or more active 33.3% or more active20% or more active

Figure 2: Probabilities of time at least the specified fraction of sources is active, using the binomial distribution

8

Possible Solution Directions:Statistical Multiplexing Gains Contd.• What are the similarities between figures 1 and 2 obtained

through simulations and analytically respectively?

– The queue delay variations with the number of aggregations (Figure 1) shares similarities with the curves obtained using the binomial distribution with different levels of active sources (Figure 2)

• For example, H=0.80 AND H=0.70 curves are very much similar in their variations with 50% or more active and 80% or more active curves respectively

9

Possible Solution Directions:Statistical Multiplexing Gains Contd.

• What inferences can be made comparing figures 1 and 2?

– They give some insights into why mean queue delay reduces when the number of self-similar aggregations increases:

• Reduction of the probability of time a specified number of sources are active

– Further the level of active sources required to have a significant impact on the queue delay determines the shape of the curve

• The level of active sources that makes a significant impact on delay is dependent on the Hurst parameter of the traffic stream

– The higher the Hurst parameter (H), the lower the level of active sources required to have a significant effect on queue delay

10

Working Through a Solution: Priority Queue with Lower Real-Time Utilisation

• How do we organize the possible solution into an implementation model?– Data packets are served at each node in a queue – Mean delay of a general queuing discipline is given by

SWD W Swhere is the average wait time in the buffer and is the average service time

– Let the instantaneous queue size, the number of packets waiting to be served at the time of the arrival of a packet be denoted by i and, the probability that the queue-size is equal to i is pi

– Then for constant service time x and maximum queue-size of N for the period of time, the mean delay is given by

(1)

11

Working Through a Solution : Priority Queue with Lower Real-Time Utilisation Contd.

xixpDN

ii

0

N

iiipxD

0

)1(

– Now consider mean delay calculated from a slowly decreasing (or nearly constant) pi function at fairly large N values, in the order of hundreds or more (as in a heavy tail distribution)

• gets extremely large as N gets largerD

(2)

(3)

12


– The above is demonstrated in the following equation

110

N

Mic

M

ii ipipxD

That is,

12

))(1(

0

MNNMpipxD c

M

ii

where pc is a constant probability value and M and N are very large. The summation term from i = M+1 to i = N has a greater impact on mean delay as M and N get larger

(4)

(5)

13


– Further (4) can be written as

where is the delay component composed of exponential decay portion of queue-size distribution

whereas is the component composed of the power-law decay portion

1 pe DDxD

(6)

Figure 3: Variation Dp with Maximum Queue

Size N for an Arbitrarily Small pc Value of

0.01 and M=100

eD

pD

14


– As seen from (5) and Figure 3, when the queue-size becomes larger the resulting mean queue delay increases exponentially

• What causes the instantaneous queue size to grow?– Hurst parameter (H)

• H depends on the traffic stream or the source

– Utilization of aggregated traffic • It depends on the number of flows aggregated and the arrival rate of

Each flow

• How controllable are H and real-time utilization level (RTF)? – H is a traffic flow attribute whereas the real time utilization level

is an attribute that the network operator has control of

• Therefore, it is possible to control the queue-size, in effect mean queue delay, even with streams of very high H with an appropriately small utilization level

15


• We have already seen the effect of H on delay– In the same way as mean delay increased with H, it also increases

with RTF; values for 70% and 80% RTF were very much higher than the value for 60% RTF

0

50

100

150

200

0 2 4 6 8 10 12 14

Number of micro f low s

Mea

n qu

eue

dela

y in

ms

H=0.80-RTF 60 H=0.82-RTF 60 H=0.80-RTF 80

H=0.82-RTF 80 Exponential H=0.82-RTF 70

H=0.80-RTF 70

Figure 5. Variation of Mean Queue Delay with RTF, H and Number of

Micro Flows

16


• What we have seen is that instantaneous queue size gets larger as the level of utilization or H get larger

– In fact, it has a dramatic increase beyond a certain H and utilization level

– For example, in Table 2, for the same H value of 0.82, 60% utilization produces a mean delay of 26ms whereas it was 180ms for 80% utilization

Table 2: Variation of Queue Size and Mean Delay with Different Traffic

Conditions

Traffic Conditions

RTF 80% RTF 80% RTF 60% Measured Parameter

H = 0.82 H = 0.80 H = 0.82 Exponential

Maximum queue-size 1810 1151 644 15

Maximum mean delay 180 78 26 3.6

17


• When real-time traffic utilization level is maintained at a lower value how can we achieve a higher overall queue efficiency?

– Use a class-based priority queuing discipline with real-time traffic treated with the highest priority at a relatively lower utilization relative to the whole queue

• The end users can be provided with statistical guarantees for their real-time data

– Non-real-time data are treated with lower priorities with no performance guarantees

18

BENEFITS OF PRIORITY QUEUEING WITH LOWER REAL-TIME UTILISATION THRESHOLD ON OPTICAL BURST SWITCHING

19

Optical Burst Switching (OBS)• What does it mean by optical burst switching?

– Vast bandwidth offered by optical wavelength-division multiplexing (WDM) technology can be merged with IP networks for cost effective data transmission

– Optical Burst Switching (OBS) is a prime candidate for switching in the optical domain using currently available technology

– OBS combines the positive features of both wavelength routing (WR) and optical packet switching (OPS)

20

OBS with PQ-LRTU

• Is there any benefit of using PQ-LRTU with OBS? If Yes, what are they?

– In this section, attempts are made to bring out the positive implications of PQ-LRTU framework, implemented in the electrical domain, to the optical domain, more specifically on OBS

– That is, we put forth that better end-to-end performance for real time traffic can be achieved when PQ-LRTU framework is merged with WDM networks using OBS

• Simulated environment and the results are shown in the next two slides

21

Suggested Environment

Figure 7: Electrical domain using PQ-LRTU framework and the Optical domain using OBS

22

Effects of Burst Assembling / Disassembling Process on Mean Delay

Variation of delay (OBS and Non-OBS) with RTF

0.00

50.00

100.00

150.00

200.00

250.00

300.00

350.00

400.00

450.00

500.00

0 2 4 6 8 10 12 14

Number of Aggregations

Dela

y (

ms)

RTF=0.2 OBS RTF=0.2 NON-OBS RTF=0.4 OBS

RTF=0.4 NON-OBS RTF=0.6 OBS RTF=0.6 NON-OBS

Figure 8: Variation of mean delay of OBS and non-OBS flows with RTF

23

Effects of Burst Assembling / Disassembling Process on Mean Delay Contd.

• What is the significance of delay variation in the previous diagram?

– The queuing phenomenon, underlying the burst assembling and disassembling process, introduced additional latency compared to their non-OBS counterparts

– This additional latency was more significant at higher RTF levels, when combined with lower aggregation levels

– In addition to the considerable additional latency in OBS traffic, at all levels of aggregations, mean delay was seen as an increasing function aggregation, beyond a certain threshold value of aggregation (See Figure 9 in the next slide)

24

Effects of Real-Time Utilization Threshold (RTF) on Burst Assembling/Disassembling Process

Variation of OBS delay with RTF

0.00

200.00

400.00

600.00

800.00

1000.00

1200.00

1400.00

0 2 4 6 8 10 12 14

Number of Aggregations

Dela

y (

ms)

RTF=0.2 RTF=0.4 RTF=0.6 RTF=0.7

Figure 9: Variation of mean queue delay of OBS flows with RTF

25

Effects of Real-Time Utilization Threshold (RTF) on Burst Assembling/Disassembling Process Contd.

• What is the significance of delay variation with RTF in OBS?

– As in the case of non-OBS traffic, mean end-to-end delay is a

decreasing function of RTF level

• The reduction being significant for lower levels of aggregation

– There is more apparent increase in delay with the level of aggregation, for lower RTF levels

– On the whole, an RTF level around 0.4, seemingly, results in somewhat uniform ideal range of delay, for all levels of aggregation

26

OBS with PQ-LRTU: Important Findings

• What are the benefits PQ-LRTU carries to optical domain, OBS in particular?

– The underlying queuing phenomenon of the burst assembling / disassembling process of OBS introduced a considerable amount of latency, generally, at all levels of aggregation

– Despite this adverse effect of burst forming process, the PQ-LRTU

framework carries its benefits to optical domain, except for one significant difference

• Unlike in non-OBS flows, where delay reduces to the extent the RTF is lowered, in OBS flows, there tend to be an optimal RTF value that results in minimum delay

27

OBS with PQ-LRTU: Important Findings Contd.

– It is also important to differentiate that additional latency incurred in OBS is not merely due to assembling / disassembling tasks, but due to their effects on changing the dynamics within the queues themselves

– The delay reduction, with the level of RTF, is significant for lower levels of aggregation, and these reductions are bound to be more significant, when larger burst sizes are used

– PQ-LRTU will accompany its other benefits, such as less starvation of lower priority packets etc. to the optical domain

28

OBS with Composite Burst Assembly Used with PQ-LRTU

• How do we overcome any limitation of PQ-LRTU on OBS, if any?

– The purpose of this section is to highlight composite burst assembly as means of overcoming adverse effects of higher burst sizes on queue delay, when aggregate self-similar traffic is present within the electrical domain

– Further, we highlight the benefits of combining a lower real-time class traffic utilization threshold with composite burst assembly for higher performance gains in aggregate, class-based traffic environment

29

Comparison of Effects of Composite andNon-Composite Burst Size on Average Delay

(a) RTF = 0.4(b) RTF = 0.6

Figure 14

30

Composite OBS with PQ-LRTU: Important Findings

• What are the benefits of using composite OBS with PQ-LRTU?

– When larger burst sizes are used in OBS, with non-composite burst assembly, a counter-productive performance effect is caused in a framework using a lower real-time class utilization threshold

– This negative impact is more significant at lower real time utilization threshold levels (RTF levels), which necessitated a tradeoff between the burst size and the RTF level, in order to maximize effectiveness

31

Composite OBS with PQ-LRTU: Important Findings Contd.

– However, the composite burst assembly mechanism removes the negative performance effects on the highest-priority, or real time, traffic class at higher burst sizes.

– As a result, the above framework can be used in conjunction with the composite burst assembly technique, without compromising the prospective performance gains for the real time traffic class

32

THANK YOU!

Any Questions?

Korea Uni Presentation

Documents