influence of file size distribution on legacy lan qos parameters

Influence of File Size Distributionon Legacy LAN QoS Parameters

Nikolaus FärberNov. 15, 2000

Outline

Network topology Qos parameter voice Traffic model and QoS parameter data PDF of file size

Uniform Log-Normal

Tradeoff QoS voice vs. data Tradeoff delay vs. loss

as before

Topology: N to N Communication

Voice received from WAN Each terminal sends/receives data to/from every other

terminal Balanced N to N communication N=16 W={1, 2, 4, 8, 16, 32, 64} = {0.1, 0.2, 0.3, 0.4, 0.5}

T2

S R

A1

AN

A2

QoS provided

WAN…

R0 = 10 Mbps

100 KByte/port, drop tail

QoS Parameter Voice

Average Voice Jitter

Reasonable quantity to predict performance of adaptive playout scheduling

More complete (but less compact) description of voice quality is possible by plotting tradeoff delay vs. loss

= | di – di+1|i=1

N1N

delay

loss

Traffic Model and QoS Par. Data Random file size Bi distributed according to fB(B) Waiting time in between file transfers:

Wi =Bi N-1

R0 R0 = 10 Mbps= load in [0,1]

time

time

B1

B2

B3

W1 W2

T1 T2

req

uest

serv

e

R = Bi

Ti

Qos parameter data: “Data goodput”:

?

File Size Distribution PB(B)

Assumed so far: Uniform (4-512 packets of 1480 byte) Literature [Barford 98, Paxson 95, Douceur 99, Arlitt 99]

File system: Log-Normal, Log-Normal Body/Pareto Tail

Network: Log-Normal, Pareto Pareto:

Log-Normal:

2

2

2

)(lnexp

2

1 )(

B

BBpB

kBBkBpB ; )( )1( (“heavy tailed”)

Workload of 1998 World Cup M. Arlitt, T. Jin, “Workload Characterization of the 1998 World

Cup Web Site”, HP Lab. Tech. Report, September 1999. http://www.hpl.hp.com/techreports/1999/HPL-1999-35R1.html Result of 1.35 billion requests during 1 month

Log-Normal

= 10.13

= 2.19

Comparison of used PDFs

file size [byte]

log2(File size)

pro

b.

pro

b.

= 3.8 105

= 2.2 105

= 3.6 103

= 1.2 104

QoS Tradeoff, PB(B) Uniform N = 16, = {0.1, 0.2, 0.3, 0.4, 0.5} x W = {1, 2, 4, 8, 16, 32, 64}

data goodput [Mbps]

avera

ge v

oic

e jit

ter

[ms]

0.3

0.1

0.2

= 0.5

14

8

16

32

W=64

2

0.4

QoS Tradeoff, PB(B) Log-Normal

N = 16, = {0.1, 0.2, 0.3, 0.4, 0.5} x W = {1, 2, 4, 8, 16, 32, 64}

data goodput [Mbps]

avera

ge v

oic

e jit

ter

[ms]

0.3

0.1

0.2

= 0.5

1

4

8

16

32

W=64

2

0.4

Uniform vs. Log-Normal PDF In general similar behavior

Average jitter decreases monotonically with window size Maximum goodput at low-medium window size (W = 4-16) High variation of goodput at low loads High variation of jitter at high loads

Longer average file size (uniform) results in reduced average voice jitter

For given scenario W=4 gives good performance at all loads Why? BxD = WxN

increase with load?

Delay vs. Loss at 10% Load

delay [ms]

loss W = {1, 2, 4, 8, 16, 32, 64}

Delay vs. Loss at 20% Load

delay [ms]

loss W = {1, 2, 4, 8, 16, 32, 64}

Delay vs. Loss at 30% Load

delay [ms]

loss W = {1, 2, 4, 8, 16, 32, 64}

Delay vs. Loss at 40% Load

delay [ms]

loss W = {1, 2, 4, 8, 16, 32, 64}

Delay vs. Loss at 50% Load

delay [ms]

loss W = {1, 2, 4, 8, 16, 32, 64}

Conclusions and Future Work Different file size distributions results in

Same general behavior Different quantitative behavior (average voice jitter)

Fixed value for window size may not be too bad

Compare Delay-Loss curves for Reduced TCP window size Adaptive playout

Further refinement of traffic model