Lecture 4: TCP and TFRC

1

TCP"Reno, NewReno, SACK

4 September 2009 CS5229, Semester 1, 2009/10

2

TCP Reno


3

0

2

4

6

8

10

12

14 cwnd ssthresh


5

new ack:"if (cwnd < sstresh)" cwnd += 1"else" cwnd += 1/cwnd


6

timeout:"retransmit 1st unacked"ssthresh = cwnd/2"cwnd = 1"


7

3rd duplicate ACK:"fast retransmission" (ie, retransmit 1st unack) "fast recovery" (details today)"ssthresh = cwnd = cwnd/2


10

TCP’s rule"

send more packets if "L + cwnd > H"

[ L .. H-1 ] are outstanding packets 4 September 2009 CS5229, Semester 1, 2009/10

11 4 September 2009 CS5229, Semester 1, 2009/10

12

Data Packets

ACK Packets (received up to this seq number)

Congestion Window

Sequence Number

Time 4 September 2009 CS5229, Semester 1, 2009/10

13

3rd dup ack:"retransmit 1st unacked"ssthresh = cwnd/2"cwnd = cwnd/2


14

X

Incorrectly clamping the congestion window too soon (without fast recovery)


15

fast recovery:"

keep the pipe occupied"


16

X



18

X

X


19

TCP Reno timeout with multiple losses in a window


20

TCP NewReno


21

X

X



25

3rd dup ack:"retransmit 1st unacked"ssthresh = cwnd/2"cwnd = cwnd/2 + 3"remember highest


26

“complete” ack:"(all are acked) "cwnd = ssthresh


27

“partial” ack:"(acknowledge n packets)"retransmit"cwnd = cwnd - n + 1


29

Note: RFC2581/RFC2582 give the accurate/gory details. Simplified version is presented here (eg. cwnd vs FlightSize, update of cwnd upon partial ACK).


30

What does a dup ACK tell us?


31

“Coarse Feedback”


32

TCP SACK


33

Use TCP header options to report

received segments.


34

SACK Blocks:"

1st block - report most recently received segments"

subsequent blocks - repeat most recent previous blocks


35

pipe: num of outstanding packets in the pipe."

send only if pipe < cwnd


36

scoreboard: which packets have been received?


37

X

X


38

3rd dup ack:"pipe = cwnd - 3"retransmit 1st unacked"ssthresh = cwnd/2"cwnd = cwnd/2 + 3


39

subsequent dup ack:"cwnd++"pipe--"

if pipe < cwnd" send packet, pipe++


40

“partial” ack:"retransmit"cwnd = cwnd - n + 1"pipe -= 2"if pipe < cwnd" send packet, pipe++


43

Idea of SACK:"

Which packet has left the network?"Where is the gap?"

Decouple when to send and what to send.


44

TCP SACK recovers faster than NewReno with multiple losses in a window.


57

Deployment


58

77%

15%

8%

NewReno

Reno

Tahoe

Feb 2004


59

70% "SACK capable


63

TCP BIC/CUBIC"Linux 2.6.x


64

Compound TCP "MS Windows Vista


65

TFRC"equation-based"

congestion control


66

Not everyone uses TCP


67

UDP:"

Media streaming"Gaming"

VoIP


68

Why not congestion controlled?"

1. UDP has low delay, no need full reliability. UDP "

is not congestion controlled.


69

Why not congestion controlled?"

2. No incentive. OTOH, there are incentives NOT to use

congestion control.


70 Sally Floyd, h+p://www.icir.org/floyd/tcp_unfriendly.html 4 September 2009 CS5229, Semester 1, 2009/10

71

“Unresponsive Flows”


72

Bad: lead to unfairness and congestion collapse.


73

Unfairness: unresponsive flows consume more bandwidth

than congestion controlled flows.


77

Congestion Collapse: wasting bandwidth by sending packets that will be dropped


78

Today: "a TCP-friendly"

unreliable protocol


79

Idea: "send at a rate that a TCP

flow would send


80

we can do the AIMD-thing"at the source, or



82

equation-based "congestion control


83

steady"fair"

responsive


84

how to determine"RTT"tRTO"p


85

p is not packet loss rate"but "

loss event rate





89

RTT can fluctuates






94

how to initialize?


95

slow start"(until loss occur)"

Tnow = min(2Tprev, 2Trecv)


96

no loss history, how?"

solve p given T, RTT



98

TFRC is now part of DCCP


Education

Lecture 4: TCP and TFRC