TCP Reno vs. Vegas › ~jst › cse › 770 › talks › georg10-14-04.pdf · Manfred Georg...

Manfred Georg October 14th 2004

1

TCP Reno vs. Vegas

Paper by:

Jeonghoon MoRichard LaVenkat AnantharamJean Walrand

INFOCOM 1999

Manfred Georg

Presented by:

Discussion Leader:

Jack Meier

Manfred Georg October 14th 2004 2

Overview

● TCP Reno● TCP Vegas● Fairness● Simulation● Conclusion

Manfred Georg October 14th 2004 3

TCP Reno

● Slow Start● Congestion Avoidance (AIMD)● Packet Loss● Window size and Network Buffering

Manfred Georg October 14th 2004 4

TCP Vegas

● Slow Start● Congestion Avoidence

– Round Trip Delay– Alpha and Beta

● Rerouting● Persistent Congestion

Manfred Georg October 14th 2004 5

Vegas Algorithm

● Expected = CWND / BaseRTT● Actual = CWND / RTT● Diff = (Expected - Actual) BaseRTT

CWNDBaseRTTRTT

ExpectedActualDiff

= 60 packets= 20 ms= 30 ms

= 3000 packets/s= 2000 packets/s= 20 packets

Manfred Georg October 14th 2004 6

Vegas Algorithm

● Expected = CWND / BaseRTT● Actual = CWND / RTT● Diff = (Expected - Actual) BaseRTT

CWND =

CWND +1

CWND – 1

CWND

if Diff < alpha

if Diff > beta

otherwise

Manfred Georg October 14th 2004 7

Vegas Algorithm

● Diff = (Expected - Actual) BaseRTT

Manfred Georg October 14th 2004 8

Rerouting Vegas

● No Explicit Feedback● Propagation Delay Changes

– Shorter Delay Detected– Longer Delay Interpreted as Congestion

● Occasionally update BaseRTT– BaseRTT = min RTT since last update– If (RTT < BaseRTT) update BaseRTT = RTT

● Cause Persistent Congestion?

Manfred Georg October 14th 2004 9

Persistent Congestion

● New Connection Increases Congestion● BaseRTT updating

– Temporarily causes congestion– Other connections back off– All connections get accurate estimate of

BaseRTT

Manfred Georg October 14th 2004 10

Fairness

● Reno vs. Reno● Vegas vs. Vegas● Reno vs. Vegas

Manfred Georg October 14th 2004 11

Reno Fairness

D. Chiu and R. Jain,Journal of Computer Networks and ISDN Systems, 1989

Manfred Georg October 14th 2004 12

Vegas Fairnessz(t) = total backlog

ei(t) = ACK received rate

qi(t) = backlog in queue

wi(t) = window size

qi(t) between and ßα

Manfred Georg October 14th 2004 13

Vegas Fairnessz(t) = total backlog

ei(t) = ACK received rate

qi(t) = backlog in queue

wi(t) = window size

wi(t) = q

i(t) + e

i(t) d

i )

ei(t) = w

i(t) / (z

i(t) + d

i)

e

1(t) + e

2(t) = 1

Manfred Georg October 14th 2004 14

Vegas Fairness

Fixed WindowStream

ß α

Manfred Georg October 14th 2004 15

Vegas Fairness

Fixed WindowStream

ß

α

Manfred Georg October 14th 2004 16

Vegas Fairness

Manfred Georg October 14th 2004 17

Vegas Fairness

80

Delay12.9ms

70

Delay18.9ms

Manfred Georg October 14th 2004 18

Vegas Fairness

110

Delay12.9ms

70

Delay36.9ms

Manfred Georg October 14th 2004 19

Vegas Fairness

160

Delay12.9ms

70

Delay66.9ms

Manfred Georg October 14th 2004 20

Vegas Fairness

400

Delay12.9ms

70

Delay186.9ms

Manfred Georg October 14th 2004 21

Simulation

● Vegas vs. Vegas● Reno vs. Reno● Reno vs. Vegas

Manfred Georg October 14th 2004 22

Vegas vs. Vegas

Manfred Georg October 14th 2004 23

Reno vs. Reno

Manfred Georg October 14th 2004 24

Reno vs. Vegas

● Vegas less than ß packets buffered– k

● Reno fills the rest (half the time)– (B-k)/2

● Vegas to Reno Throughput– 2k/(B-k)

Manfred Georg October 14th 2004 25

Reno vs. Vegas

Manfred Georg October 14th 2004 26

Reno vs. Vegas

Manfred Georg October 14th 2004 27

Conclusion

● Rerouting● Persistent Congestion● Vegas vs. Vegas Fairness● Reno vs. Reno Fairness● Vegas and Reno do not cooperate

Manfred Georg October 14th 2004 28

Thank You

'The time has come,' Jack Meier said,'To speak of many thing...'