Congestion Control 2 Talks

TCP Congestion Control:
Algorithms and Analysis

Simon S. Lam
Department of Computer Sciences
The University of Texas at Austin
Littles Law
Average
g population
p p
= (average delay) x
(throughput)
average delay 1
N
delayi
i1
where N is number of departures
throughput N/T
where T is duration of observation
average population (to be defined)
Try homework problem at

http://www.cs.utexas.edu/users/lam/cs356/homework/hw2.html
TCP Congestion Control (Simon Lam)
Number in system n(t)
Time t
average population 1 T0 n(t)dt
T
Sliding Window Protocol

Consider an infinite array, Source, at the
sender, and an infinite array, Sink, at the

receiver.
send window
Source:
P1
Sender
a1 a
acknowledged
P2
Receiver
Sink
0
unacknowledged
next expected
received
:
1
s1 s
r + RW 1
r
delivered
RW receive window size
SW send window size (s - a SW)
receive window
Sliding Windows in Action

Data unit r has just been received by P2
Receive window slides forward
P2
P sends
d cumulative
l
ack
k with
h sequence
number it expects to receive next (r+3)

send window
Source:
P1
Sender
a1 a
acknowledged
s1 s
unacknowledged
r+3
next expected
Sink:
P2
Receiver
r + RW 1
r
delivered
receive window
Sliding Windows in Action

P1 has just received cumulative ack with
r+3 as next expected sequence number
Send window slides forward

send window
Source:
P1
Sender
a1 a
s1 s
acknowledged
next expected
Sink:
P2
Receiver
r + RW 1
r
delivered
receive window
Window Flow Control

RTT
Source
1 2
1 2
time
ACKs
data
Destination
1 2
1 2
time
~ W packets per RTT when no loss

Lost packet detected by missing ACK
(note: timeout value TO > RTT)
Throughput (send rate)

Limit the number of unacked transmitted
packets in the network to window size W
Throughput
W
RTT
packets/sec
W MSS
RTT
bytes/sec
Where did we apply Littles Law?

Answer: Consider send buffer
Clarifications
Average number in the send buffer is typically
less than W unless packet arrival rate to send

buffer is infinite -> previous formula provides a
throughput upper bound
If each packet may be lost with rate p, then the
average delay is
(1 p) RTT p TO
Since TO > RTT, actual throughput is smaller.

With loss, g
goodput
p is
(1 p ) throughput
th
h t
Note: in some papers and other context (e.g., random

access protocols), goodput is called throughput. To avoid
confusion, throughput is called send rate
Effect of Congestion
W too big for each of many flows -> congestion
Packet loss -> transmissions on links prior to packet
loss are wasted

Congestion collapse due too many retransmissions
and too much waste
October 1986, Internet had its first congestion
collapse
goodput
load
10
TCP Window Control

Receiver flow control
Avoid overloading receiver
rwnd:
receiver (advertised) window
Receiver sends rwnd to sender
Network congestion control
Sender tries to avoid overloading network
It infers available network capacity from loss
i di ti
indications
cwnd:
congestion window
Sender sets W = min (cwnd, rwnd)
11
Receiver Flow Control

Receiver advertises rwnd with each packet it
sends
Size of rwnd indicates available space in
receive buffer
decreased when data is received from IP layer and

ackd
increased when data is consumed by application
process
12
Network Congestion Control

Sender calculates cwnd from indications of
network congestion
Congestion indications
timeout (loss)
dupACK (loss likely)
queueing delay
mark (needs ECN)
TCP algorithms to calculate cwnd

Tahoe, Reno, Vegas,
Link algorithms:
RED, REM
13
TCP & AQM

pl((t))
xi(t)
Congestion measures pl(t) for distributed feedback control of xi(t)

loss and dupACK (DropTail)
queueing delay (Vegas)
with the help of active queue management (AQM)
queue length (RED)
price (REM)
14
TCP Congestion Control

Tahoe (Jacobson 1988)
Slow Start
Congestion Avoidance
Fast
F t Retransmit
R t
it
Reno (Jacobson 1990)
Fast Recovery
Its variants: NewReno, SACK
Vegas (Brakmo & Peterson 1994)
New Congestion Avoidance
AQM
RED
E (Floyd
(Fl d & Jacobson
b
1993)
1
)
Probabilistic marking or dropping
REM (Athuraliya & Low 2000)

Clear buffer, match rate
Others
15
Slow Start
Start with cwnd = 1
On each successful ACK, increment cwnd
cwnd cwnd + 1
Exponential growth of cwnd
each RTT: cwnd 2 x cwnd
Enter CA when cwnd >= ssthresh

For initial slow start,, ssthresh is set to a very
y large
g
value (e.g., 65 Kbytes)
Note: for clarity, cwnd, rwnd, and ssthresh are

counted in packets (segments) rather than in bytes
16
Slow Start
sender
receiver
cwnd
1
1 RTT
2
data packet
ACK
3
4
5
6
7
8
cwnd cwnd + 1 (for each ACK)TCP Congestion Control (Simon Lam)
17
Congestion Avoidance
sender
receiver
cwnd
1
cwnd ssthresh
data packet
ACK
CA starts when
1 RTT
On each successful
ACK:
cwnd cwnd + 1/cwnd
Linear growth of cwnd

4
each RTT:
cwnd cwnd + 1
18
Packet Loss
Assumption: loss indicates congestion
Packet
P k t loss
l
detected
d t t db
by
Retransmission timeout (RTO timer)

Duplicate ACKs (at least 3)
Packets
Acknowledgements
19
Fast Retransmit
A timeout is quite long (> RTT)
Upon
p receiving
g 3 dupACKs,
p
, immediately
y
retransmit without waiting for timeout
Adjusts ssthresh
ssthresh max(flightsize/2, 2)
where
h
flightsize
fli htsiz is numb
number of
f outstanding
utst ndin p
packets,
ck ts
which may be less than W = min(rwnd, cwnd)
Enter Slow Start (cwnd = 1)
20
10
TCP Tahoe
(Jacobson 1988)
cwnd
SS
time
(in RTTs)
CA
SS: Slow Start

CA: Congestion Avoidance
21
Successive Timeouts
When there is another timeout, double the
timeout value
Keep doing so for each additional lossretransmission
Exponential backoff up to
max timeout value equal
to 64 times initial timeout
value
Note: red line in figure denotes a loss indication

22
11
Summary: Tahoe
Basic ideas
Probe network for spare capacity during SS and
CA and increase send rate
Drastically reduce rate on congestion indication
Self-clocking
Error recovery by retransmission
Round trip time estimation (to get TO value)
for every ACK {
if (W < ssthresh) then W++
(SS)
else W += 1/W
(CA)
}
for every loss indication {
ssthresh = W/2
W = 1
}
TCP Tahoe
23
(Jacobson 1988)
cwnd
SS
time
CA
SS: Slow Start

24
12
TCP Reno
(Jacobson 1990)
cwnd
SS
time
CA
SS: Slow Start

Fast retransmission/fast recovery

25
TCP Reno (another scenario)

TO
cwnd
3 dupACKs
halved
Initial slow start
Slow start until cwnd

reaches ssthresh
26
13
Fast recovery (in detail)

Idea: each dupACK represents a packet
successfully received.
received Therefore,
Therefore no need for
very drastic action
Enter FR/FR after 3 dupACKs
Set ssthresh max(flightsize/2, 2)

Retransmit lost packet
Set cwnd ssthresh + #dupACKs (window inflation)
Wait till W=min(rwnd
W=min(rwnd, cwnd) is large enough; transmit
new packet(s)
On non-dup ACK (1 RTT later), set cwnd ssthresh
(window deflation)
Enter CA
27
Example: FR/FR
S 1 2 3 4 5 6 7 8
9 10 11
time
Exit FR/FR
cwnd 8
ssthresh
7
4
time
0 0 0 0 0 0 0
9
4
11
4
4
4
Above scenario: Packet 1 is lost, p

packets 2, 3, and
4 are received;
i d dupACKs
d ACK with
i h seq. no. 0 returned
d
Fast retransmit
Retransmit on 3 dupACKs
Fast recovery
Inflate window such that new packets 9, 10, and 11 can be
sent while repairing loss
28
14
Summary: Reno
Basic ideas
Fast recovery avoids slow start
dupACKs: fast retransmit + fast recovery
Timeout: fast retransmit + slow start
dupACKs
congestion
avoidance
FR/FR
timeout
slow start
retransmit
29
AIMD in steady state

additive increase:
increase cwnd by 1
MSS every RTT in the
absence
bs
of
f any lloss
ss
event
multiplicative decrease:
cut cwnd in half after
3 dupACKs
congestion
window
24 Kbytes
16 Kbytes
8 Kbytes
Long-lived TCP connection
time
30
15
TCP throughput (send rate)

We derived the approximate formula
throughput =
W
packets/sec
RTT
W changes with the arrival of each
congestion indication
To calculate
l l
(average)
(
) send
d rate, we need
d
the average value of W
Q: W is a function of what parameter?
31
First approximation
M. Mathis, et al., The Macroscopic Behavior of the TCP Congestion

Avoidance Algorithm,ACM Computer Communicatons Review, 27(3), 1997.
No slow-start, no timeout, long-lived TCP
connection
Independent identically distributed periods
Each packet may be lost with probability p
32
16
Geometric Distribution
Ave. no. of transmissions to get first loss

n
ibi i (1 p)i1 p
i 1
i 1
p i (1 p )i 1
i 1
d
d
(1 p )i p
dp i 1
dp
d
1
1
p
p 2
dp 1 1 p
p
1/ p
p
(1 p)
i 0
Similarly, ave. no. of transmissions to get first success is

1/(1-p)
33
First approximation (cont.)

Average number of
packets delivered in
one period (area under
one saw-tooth)
2
W 1 W 3 2
W
2 2 2 8
Average number of
packets
k
sent per period
i d
(incl. loss at the end) is
1/p
Equating the two and
solving for W, we get
8
3p
send rate (in packets/sec)

no. of packets/period
time per period
1/ p
1
2 RTT
RTT
3p
3 2
W
8
W
RTT
2
3
2p
34
17
TCP ACK generation
[RFC 1122, RFC 2581]
Event at Receiver
TCP Receiver action
Arrival of in-order segment with

expected seq #. All data up to
expected seq # already ACKed
Delayed ACK. Wait up to 500ms

for next segment. If no next segment,
send ACK
Arrival of in-order segment with

expected seq #. One other
segment has ACK pending
Immediately send single cumulative

ACK, ACKing both in-order segments
Arrival of out
out-of-order
of order segment
higher-than-expect seq. # .
Gap detected
Immediately send duplicate ACK

ACK,
indicating seq. # of next expected byte
Arrival of segment that

partially or completely fills gap
Immediate send ACK, provided that

segment starts at lower end of gap
35
Receiver implements Delayed ACKs

Receiver sends one ACK for every two packets
received -> each saw-tooth is WxRTT wide

-> area under a saw-tooth
h is
3W 2
Send rate is
1
RTT
One ACK for every
i
is
1
RTT
3
4p
b packets received -> send rate
3
2bp
36
18
Modeling
M
d li TCP Throughput:
Th
h t A Si
Simple
l
Model and its Empirical Validation,
Proc. ACM SIGCOMM, 1998
Jitendra Padhye, Victor Firoiu,
D Towsley,
Don
T
l
and
d Jim
Ji Kurose
K
Motivation
Previous formulas not so accurate when
loss rates are high

TCP traces show that there are more loss
indications due to timeouts (TO) than due
to triple dupACKs (TD)
38
19
Objectives
More accurate steady-state throughput
formula as a function of loss rate and RTT

by also accounting for TO behavior of a
TCP connection
Formula applicable over a wider range of
loss rates
Explicit statements of assumptions and
approximations
i ti
used
d iin d
derivation
i ti of
f
throughput formula
Formula to include the impact of a small
rwnd
39
Many assumptions and

approximations
A1. TCP sender is saturated, i.e., source
application process always has a packet to

send when send window has space available
i.e., bulk transfer application
A2. Slow Start not modeled

A3. Time to send all packets in a window is
smaller than RTT
i.e.,transmission rate is not too low
40
20
A3. Time to send W packets is

less than RTT
ACK reception
marks the end of

current round and
beginning of next
round.
Start of round
time
Approximation: For
b > 1, ACK is not
received
immediately after
one RTT, but it is so
assumed in the
analysis
End of round
space
41
AIMD evolution of Window Size over time
Each TD period is ended by a TD loss indication.

TDPi p
period has duration Ai rounds
A4. Duration of a round (RTT) is independent of
window size
approximation (poor for a slow line)
A5. No window inflation in Fast Recovery

approximation
42
21
Markov regenerative assumption

For the
i-th TD period, Wi is window size at
the end of the period, Yi is the number of

packets sent in the period
A6. Assume {Wi} to be a Markov
regenerative process with rewards {Yi}
Given A6, the steady-state TCP throughput
is
B lim Bt lim
t
N t E[Yi ] E[Y ]
t
E[ Ai ] E[ A]
43
Consider i-th TD period
when ACK of
last packet is
received
b packets (b = 2 in above
figure) -> linear increase has a slope of 1/b packet per
RTT
Number of rounds is Xi +1
One ACK after receiving
i is the first packet lost in i-th TD period
44
22
Loss assumptions
A7. Losses in different rounds are
independent
approximation
A8. Losses within the same round are
correlated as follows: If a packet is lost,

all remaining packets transmitted until the
end of that round are also lost
approximation
i ti bursty
b t lloss b
behavior
h i b
butt only
l
within the same round
all lost packets in the same round are counted
as a single loss indication when estimating p
45
AIMD throughput derivation (1)
E[ ] 1/ p
E[r] RTT
E[Y ] E[ ] E[W ] 1
From Wi
1
1 E[W ]
p
Wi1 Xi
, we have
b
2
seq. no. of first loss

r round trip time
Y no. of p
packets sent
W window size
X no. of rounds
A time duration of a period
b
E[ X ] E[W ]
2
b
E[ A] (E[ X ] 1)E[r] ( E[W ] 1)RTT
2
1
1 E[W ]
E[Y ]
p
send rate B
E[ A] ( b E[W ] 1)RTT
2
<- from A4 that round trip

times are independent of Wi
46
23
AIMD throughput derivation (2)

Another way to
compute E[Y]
X i / b 1
Wi 1
k )b i
2
k 0
XW
X X
i i 1 i ( i 1) i
2
2 b
X
W
i (Wi i 1 1) i
2
2
W
Let E[ ] be E[ ] and we have
2
E[ X ]
E[W ]
E[Y ]
( E[W ]
1) E[ ]
2
2
bE[W ]
E[W ]
W
( E[W ]
1) E[ ]
4
2
2
Yi
<- A9. Assume that

{Xi} and {Wi} are
mutually
independent i.i.d.
sequences of
random variables
47
AIMD throughput (3)

Equate the two previous formulas for E[Y]. Solve the quadratic
equation with E[W] as the only unknown
1
1 E [W ]
p
b E [W ]
E [W ]
W
( E [W ]
1) E [ ]
4
2
2
E [Y ]
E [W ]
2b
3b
se n d ra te
8(1 p )
2b 2
)
(
3b p
3b
1 p
E [W ]
p
B( p)
E [ A]
48
24
AIMD throughput (4)

To get a simple formula, collect terms that are o(1/sqrt(p))
E[W ]
8
o(1/ p )
3bp
b
2b
E[ X ] E[W ]
o(1/ p )
2
3p
send rate B( p)
1/ p o(1/ p)
1
3
o(1/ p )
2b
RTT 2bp
o(1/ p )
RTT
3p
49
AIMD with TO
Let ni denote the number of TD periods within a
cycle ending in i-th

th TO period, Ri denote no. of
retransmissions in i-th TO period
A10. {ni } form an i.i.d. sequence, independent of
{Yij} and {Aij}
50
25
Throughput of AIMD with TO (1)

E[ M ] E[n]E[Y ] E[ R ]
Assumption of
Markov
regenerative
process again.
E[ S ] E[n]E[ A] E[ Z TO ]
E[ M ]
E[n]E[Y ] E[ R]
send rate B
E[ S ] E[n]E[ A] E[ Z TO ]
E[Y ] Q E[ R]
B
E[ A] Q E[ Z TO ]
1
where Q
<- Probability
P b bilit th
thatt a
E[n]
given loss
1
E[ R]
indication is a TO
1 p
with Q and E[ Z TO ] to be determined
51
Approximate solution for Q
A given loss indication is a TO is the union of two

events Two or less acked packets in penultimate
round or two or less acked packets in final round
52
26
Approximate solution for Q (cont.)

<- penultimate round of w
packets, first k packets
ackd given there is a loss
(1 p ) k p
A( w, k )
1 (1 p ) w
C (k , m ) (1 p ) m p,
m k 1
C (k , m ) (1 p ) m ,
mk
( w) 1
Q
<- for last round, k packets

sent, m packets ackd in
sequence
<- at most 2 dupACKs

p
2
w
2
<- probability of fewer than 3
A( w, k ) A( w, k ) C ( k , m) packets sent successfully
k 0
k 3
m 0
in penultimate round or
if w 4
less than 3 acks in last
round
if w 3
53
Approximate solution for Q (cont.)
Q is
( w)]
E[Q
But we dont know the probability distribution of Wi

Approximation
3
3bp
Q Q ( E[W ]) min(1,
E[W ]
) min(1,3
54
27

P[ R k ] p k 1 (1 p )
Lk (2 1)TO
k
for k 1, 2,...
for k 6
(63 64(k 6))TO
for k 7
<- duration of k
TOs in a row
1 p 2 p 2 4 p 3 8 p 4 16 p 5 32 p 6
1 p
f ( p)
<- approximation
TO
T0 (1 32 p 2 )
1 p
E[Y ] Q E[ R ]
sendd rate B ( p )
E[ A] Q E[ Z TO ]
1 p
( E[W ]) 1
E[W ] Q
p
1 p
B( p)
( E[W ])T f ( p )
RTT ( E[ X ] 1) Q
O
1 p TCP Congestion Control (Simon Lam)
E[ Z TO ] TO
55

1 p
( E[W ]) 1
E[W ] Q
<- Eq. (27)
p
1 p
B( p)
more accurate
( E[W ])T f ( p)
RTT ( E[ X ] 1) Q
version of
O
1 p
throughput
formula
1/ p
2b
3bp
2
RTT
min 1,3
(1 32 p )T0
3
p
8
1
<-Eq. (29)
2bp
3bp
2
most wellRTT
min 1,3
p(1 32 p )T0
known version
3
8
of throughput
formula
56
28
Impact of receivers rwnd limitation
Full model Eq. (31)
Compute E[W ]. If E[W ] Wmax , use Eq. (27):
1 p
( E[W ]) 1
E[W ] Q
p
1 p
B( p)
if E[W] <Wmax,
( E[W ])T f ( p)
RTT ( E[ X ] 1) Q
O
1 p
1 p
(W ) 1
Wmax Q
otherwise, use Wmax for
max
p
1 p
E[W] and recompute
B( p)
b
1 p
f ( p)
E[X]
2) Q(Wmax )TO
RTT ( Wmax
8
1 p
pWmax
(derivation omitted)
57
Impact of receivers rwnd

limitationapproximate model
Use the well-known Eq. (29) from before,
B( p ) min(
Wmax
,
RTT
1
2bp
3bp
2
RTT
min 1,3
p (1 32 p )T0
8
3
which is referred to as Eq. (32)

58
29
Summary data from traces (1 hour)

Saturated TCP
sender
p computed
from dividing
total no. of loss
indications by
total number of
packets sent
RTT and TO
values are
averaged over
entire 1-hour
trace
59
Summary data from 100s traces
Each row represents results of 100 traces each of
100 seconds in duration for same S-D pair

Totals are cumulative over 100 traces
RTT and TO are average values over 100 traces for
same S-D pair
60
30
Experimental comparison (1)
Each point represents number of packets in 100s interval of trace

T0 ~ single TO, T1 ~ at least 1 double TO in trace, etc.
TD Only is analytic model by Mathis et al.
Note: Wmax is only 6 in Figure 7
61
Wmax = 33
Wmax=44
62
31
Wmax=8
Wmax=48
63
Accuracy of approximate model
Figure 18: manic to spiff, with predictions by both full and

approximate models
(Wmax=32)
64
32
Average errors
ave. error
N predicted N observed
N observed
no. of observations
observations
65
Conclusions
A much more rigorous analysis than the one by
Mathis et al.
Numerous assumptions and approximations used
but (almost) all of them are explicitly stated
Large amount of experimental measurements on
the Internet to validate accuracy of the full model
(less for the approximate model)
Throughput formula accounts for loss indications
due to TO as well as rwnd restriction
Using
U
i the
h formula
f
l requires
i
accurate measurements of
f
loss rate and RTT values (which could be tricky)
For TCP Reno and drop-tail router
Accuracy (like beauty) is in the eye of the
beholder. What do you think?
66
33
TCP Throughput limited by loss rate

TCP average throughput (approximate) in
terms of loss rate, L:

1 22 MSS
1.22
RTT p
Example: 1500-byte segments, 100ms RTT,
to get 10 Gbps throughput, loss rate needs

to be very low
10
p = 2x10
2 10-10
New version of TCP needed for connections
with high-delay bandwidth product
addressed in paper by Katabis et al
67
68
The End
34

Congestion Control 2 Talks

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Congestion Control 2 Talks

Hochgeladen von

Copyright:

Verfügbare Formate

TCP Congestion Control:

Algorithms and Analysis

where N is number of departures

where T is duration of observation

average population (to be defined)

Try homework problem at

TCP Congestion Control (Simon Lam)

Number in system n(t)

TCP Congestion Control (Simon Lam)

Sliding Window Protocol

sender, and an infinite array, Sink, at the

TCP Congestion Control (Simon Lam)

Sliding Windows in Action

number it expects to receive next (r+3)

Sliding Windows in Action

r+3 as next expected sequence number

Send window slides forward

TCP Congestion Control (Simon Lam)

Window Flow Control

~ W packets per RTT when no loss

(note: timeout value TO > RTT)

TCP Congestion Control (Simon Lam)

Throughput (send rate)

packets in the network to window size W

Where did we apply Littles Law?

less than W unless packet arrival rate to send

If each packet may be lost with rate p, then the

Since TO > RTT, actual throughput is smaller.

Note: in some papers and other context (e.g., random

loss are wasted

TCP Congestion Control (Simon Lam)

TCP Window Control

Receiver Flow Control

decreased when data is received from IP layer and

TCP Congestion Control (Simon Lam)

Network Congestion Control

TCP algorithms to calculate cwnd

TCP Congestion Control (Simon Lam)

TCP & AQM

Congestion measures pl(t) for distributed feedback control of xi(t)

TCP Congestion Control

REM (Athuraliya & Low 2000)

TCP Congestion Control (Simon Lam)

Exponential growth of cwnd

each RTT: cwnd 2 x cwnd

Enter CA when cwnd >= ssthresh

value (e.g., 65 Kbytes)

Note: for clarity, cwnd, rwnd, and ssthresh are

cwnd cwnd + 1 (for each ACK)TCP Congestion Control (Simon Lam)

Linear growth of cwnd

Retransmission timeout (RTO timer)

TCP Congestion Control (Simon Lam)

retransmit without waiting for timeout

SS: Slow Start

Note: red line in figure denotes a loss indication

SS: Slow Start

SS: Slow Start

Fast retransmission/fast recovery

TCP Reno (another scenario)

Slow start until cwnd

TCP Congestion Control (Simon Lam)

Fast recovery (in detail)

Set ssthresh max(flightsize/2, 2)

Above scenario: Packet 1 is lost, p