The effect of router buffer size on the TCP performance K.E. Avrachenkov*, U.Ayesta**, E.Altman*,...

13
The effect of router buffer size on the TCP performance K.E. Avrachenkov*, U.Ayesta**, E.Altman*, P.Nain*,C.Barakat* *INRIA - Sophia Antipolis, France ** FT R&D - Sophia Antipolis, France TELECOMMUNICATION NETWORKS AND TELETRAFFIC THEORY Thursday, January 31, 2002 Saint Petersburg, Russia

Transcript of The effect of router buffer size on the TCP performance K.E. Avrachenkov*, U.Ayesta**, E.Altman*,...

The effect of router buffer size on the TCP performance

K.E. Avrachenkov*, U.Ayesta**, E.Altman*,

P.Nain*,C.Barakat*

*INRIA - Sophia Antipolis, France** FT R&D - Sophia Antipolis, France

TELECOMMUNICATION NETWORKS AND TELETRAFFIC THEORY

Thursday, January 31, 2002

Saint Petersburg, Russia

2

Is there an optimal value for the buffer size in IP routers?

Outline

Persistent TCP connections:Average sending rate and goodput.

Short TCP connections:Latency

TCP IPpacket loss prob

sending rate

delay propagation

Fixed Point Approach.

3

Simulated Scenario

Persistent Connections: 100 connections.

100Mbps, 2ms

10Mbps, 1ms

10Mbps, 11ms

Short Transfers: New sessions arrive according to a Poisson distribution.

1

N

4

Persistent TCP Connections: Fixed Point (I)

11

1

KKp

K

K

ii KpC

MSSdRTT

11

1*2

N

iii RTTpT

C 1

,1

For the formula for see for example:

J.Padhye, V.Firoiu,D. Towsley, and J. Kurose, « Modeling TCP Throughput: a Simple Model and its Empirical Validation   », in Proceedings of ACM SIGCOMM ’1998 conference.

),( iii RTTpTT

Prop Delay

Mean queueing delay

5

Persistent TCP Connections: Fixed Point (II)

N

iK

K

iKK

i KC

MSSdT

C 11 11

1,

1

11

Solved

iteratively

Substituting we

get the values for

p,RTTi and Ti.

6

Persistent TCP connections: Result (I)

Buffer size (packets)

Rate (bps)

7

Persistent TCP connections: Result (II)

Buffer size (packets)

Rate (bps)

8

Short TCP Transfers: Fixed Point (I)

PlinkC

N

i isizedocE

11

_

11

1

K

KP

Solved

iteratively Ki

Ki

i

1

1 1

01

0

9

Short TCP Transfers: Fixed Point (II)

*) B. Sikdar et. al. An integrated model for the latency and steady-state throughput of TCP connections. Perf .Eval. 01

*

maxmax

57.157.114

31020log4,log

mWWp

RTTpmpmRTTpfmRTTmL

.10

1

1

164232.2,

33

32

RTT

pN

RTT

pppRTTpf

K

K

ii KpC

MSSdRTT

11

1*2

Substituting we get the values for RTTi and Latency.

10

Short TCP Transfers: Result

Buffer size (packets)

Latency (bps)

11

Conclusions

Persistent TCP connections has poor sending rate and goodput in the cases of small and large buffer sizes.

Consequently there is an optimal value of the buffer size.

Short TCP transfers. There seems not to be an optimal value for the IP router buffer size. The larger the buffer size, the better.

Fairness improves with increasing buffer size.

Connections with large propagation delays get more bandwith and those with short propagation delays less.

12

Future work

More sophisticated network topologies with several bottleneck links.

Formula of Sikdar et al. Does not work well in the case of large number of packet losses.

13

Formula for the Sending rate

,,

1

,2

18

,1

,

,

1

,1

2

,1

,

0max

maxmax

maxmax

max0

otherwise

p

TpFWpQ

pWp

Wb

RTT

WpQWp

p

MSSRTTpT

WpWif

p

TpFpWpQpW

bRTT

pWpQpWp

p

MSSRTTpT

ii

iij

i

ii

iii

ii

ii

iii

,32168421

,11

111111,1min,

,9/13/123/2

65432

333

pppppppF

p

pppwpQ

pppW

w

w