Embed Size (px)
Citation preview
22620
1
COMP 3825 Networking and Information Assurance
TCP Congestion Control
COMP 3825 - Kan Yang 1
1
Learning Objectives
1 principles of congestion control
2 TCP congestion control
COMP 3825 - Kan Yang 2
2
Principles of congestion control
congestionbull informally ldquotoo many sources sending too much data too fast for
network to handlerdquobull different from flow controlbull manifestations
ndash lost packets (buffer overflow at routers)ndash long delays (queueing in router buffers)
bull a top-10 problem
COMP 3825 - Kan Yang 3
3
Causescosts of congestion scenario 1sect two senders two receiverssect one router infinite buffers sect output link capacity Rsect no retransmission
sect maximum per-connection throughput R2
unlimited shared output link buffers
Host A
original data lin
Host B
throughput lout
R2
R2
l out
lin R2
dela
y
linv large delays as arrival rate lin
approaches capacityCOMP 3825 - Kan Yang 4
4
22620
2
Causescosts of congestion scenario 2
sect one router finite buffers sect sender retransmission of timed-out packet
bull application-layer input = application-layer output 120582 = 120582ampbull transport-layer input includes retransmissions 120582 ge 120582amp
finite shared output link buffers
Host A
lin original data
Host B
loutlin original data plusretransmitted data
COMP 3825 - Kan Yang 5
5
idealization perfect knowledgesect sender sends only when router
buffers available
finite shared output link buffers
lin original dataloutlin original data plus
retransmitted data
copy
free buffer space
R2
R2
l out
lin
Causescosts of congestion scenario 2
Host B
A
COMP 3825 - Kan Yang 6
6
Causescosts of congestion scenario 2
Idealization packets can be lost dropped at router due to full bufferssect sender only resends if packet known to be lost
lin original dataloutlin original data plus
retransmitted data
copy
no buffer spaceA
Host BCOMP 3825 - Kan Yang 7
7
Causescosts of congestion scenario 2
Idealization packets can be lost dropped at router due to full bufferssect sender only resends if packet known to be lost
lin original dataloutlin original data plus
retransmitted data
free buffer spaceA
Host BCOMP 3825 - Kan Yang 8
8
22620
3
Causescosts of congestion scenario 2
Realistic duplicatessect packets can be lost dropped at router
due to full bufferssect sender times out prematurely sending
two copies both of which are delivered
A
linloutlincopy
free buffer space
timeout
R2
R2lin
lout
when sending at R2 some packets are retransmissions including duplicated that are delivered
Host BCOMP 3825 - Kan Yang 9
9
Causescosts of congestion scenario 2Realistic duplicatessect packets can be lost dropped at router
due to full bufferssect sender times out prematurely sending
two copies both of which are delivered
R2
R2lin
lout
when sending at R2 some packets are retransmissions including duplicated that are delivered
ldquocostsrdquo of congestionsect more work (retrans) for given ldquogoodputrdquo (application-level throughput)sect unneeded retransmissions link carries multiple copies of pkt
bull decreasing goodput
COMP 3825 - Kan Yang 10
10
Causescosts of congestion scenario 3
sect four senderssect multihop pathssect timeoutretransmit
Q what happens as lin and linrsquo increase
finite shared output link buffers
Host A lout Host B
Host CHost D
lin original datalin original data plus
retransmitted data
A as red linrsquo increases all arriving blue pktsat upper queue are dropped blue throughput g 0
COMP 3825 - Kan Yang 11
11
another ldquocostrdquo of congestionsect when packet dropped any ldquoupstream transmission
capacityrdquo used for that packet was wasted
Causescosts of congestion scenario 3
C2
C2
l out
linrsquo
COMP 3825 - Kan Yang 12
12
22620
4
Learning Objectives
1 principles of congestion control
2 TCP congestion control
COMP 3825 - Kan Yang 13
13
TCP congestion control additive increase multiplicative decrease
sect approach sender increases transmission rate (window size) probing for usable bandwidth until loss occursbull additive increase increase cwnd (congestion window) by 1
MSS (maximum segment size) every RTT until loss detectedbull multiplicative decrease cut cwnd in half after loss
cwnd
TCP
send
er
cong
estio
n wi
ndow
size
AIMD saw toothbehavior probing
for bandwidth
additively increase window size helliphellip until loss occurs (then cut window in half)
timeCOMP 3825 - Kan Yang 14
14
TCP Congestion Control details
sect sender limits transmission
sect cwnd is dynamic function of perceived network congestion
TCP sending ratesect roughly send cwnd bytes wait
RTT for ACKS then send more bytes
last byteACKed sent not-
yet ACKed(ldquoin-flightrdquo)
last byte sent
cwnd
LastByteSent - LastByteAcked lt cwnd
sender sequence number space
rate ~~cwndRTT
bytessec
COMP 3825 - Kan Yang 15
15
TCP Slow Start sect when connection begins increase
rate exponentially until first loss eventbull initially cwnd = 1 MSSbull double cwnd every RTTbull done by incrementing cwnd for
every ACK received
sect summary initial rate is slow but ramps up exponentially fast
Host A
one segment
RTT
Host B
time
two segments
four segments
COMP 3825 - Kan Yang 16
16
22620
5
TCP detecting reacting to losssect loss indicated by timeout
bull cwnd set to 1 MSS bull window then grows exponentially (as in slow start) to threshold
then grows linearly
sect loss indicated by 3 duplicate ACKs TCP RENObull dup ACKs indicate network capable of delivering some segments bull cwnd is cut in half window then grows linearly
sect TCP Tahoe always sets cwnd to 1 (timeout or 3 duplicate acks)
COMP 3825 - Kan Yang 17
17
Fast RecoveryQ when should the exponential increase switch to linear A when cwnd gets to 12 of its value before timeout
Implementationsect variable ssthresh
(slow start threshold)sect on loss event ssthresh
is set to 12 of cwnd just before loss event
Check out the online interactive exercises for more examples httpgaiacsumassedukurose_rossinteractive
COMP 3825 - Kan Yang 18
18
TCP throughputbull avg TCP thruput as function of window size RTT
ndash ignore slow start assume always data to send
bull W window size (measured in bytes) where loss occursndash avg window size ( in-flight bytes) is frac34 Wndash avg thruput is 34W per RTT
W
W2
avg TCP thruput = 34
WRTT bytessec
COMP 3825 - Kan Yang 20
20
TCP Futures TCP over ldquolong fat pipesrdquo
bull example 1500 byte segments 100ms RTT want 10 Gbps throughputbull requires W = 83333 in-flight segmentsbull throughput in terms of segment loss probability L [Mathis 1997]
to achieve 10 Gbps throughput need a loss rate of L = 210-10 ndash a very small loss rate
bull new versions of TCP for high-speed
TCP throughput = 122 MSSRTT L
COMP 3825 - Kan Yang 21
21
22620
6
TCP Fairnessfairness goal if K TCP sessions share same bottleneck link of bandwidth R
each should have average rate of RK
TCP connection 1
bottleneckrouter
capacity RTCP connection 2
COMP 3825 - Kan Yang 22
22
Why is TCP fairtwo competing sessionssect additive increase gives slope of 1 as throughout increasessect multiplicative decrease decreases throughput proportionally
R
R
equal bandwidth share
Connection 1 throughput
Con
nect
ion
2 th
roug
hput
congestion avoidance additive increaseloss decrease window by factor of 2
congestion avoidance additive increaseloss decrease window by factor of 2
COMP 3825 - Kan Yang 23
23
Fairness (more)Fairness and UDPsect multimedia apps often do not use TCP
bull do not want rate throttled by congestion controlsect instead use UDP
bull send audiovideo at constant rate tolerate packet loss
Fairness parallel TCP connectionssect application can open multiple parallel connections between two hostssect web browsers do this sect eg link of rate R with 9 existing connections
bull new app asks for 1 TCP gets rate R10bull new app asks for 11 TCPs gets R2
COMP 3825 - Kan Yang 24
24
Explicit Congestion Notification (ECN)
network-assisted congestion controlsect two bits in IP header (ToS field) marked by network router to indicate congestionsect congestion indication carried to receiving hostsect receiver (seeing congestion indication in IP datagram) ) sets ECE (ECN-Echo) bit on
receiver-to-sender ACK segment to notify sender of congestion
source
applicationtransportnetwork
linkphysical
destination
applicationtransportnetwork
linkphysical
ECN=00 ECN=11
ECE=1
IP datagram
TCP ACK segment
COMP 3825 - Kan Yang 25
25
22620
2
Causescosts of congestion scenario 2
sect one router finite buffers sect sender retransmission of timed-out packet
bull application-layer input = application-layer output 120582 = 120582ampbull transport-layer input includes retransmissions 120582 ge 120582amp
finite shared output link buffers
Host A
lin original data
Host B
loutlin original data plusretransmitted data
COMP 3825 - Kan Yang 5
5
idealization perfect knowledgesect sender sends only when router
buffers available
finite shared output link buffers
lin original dataloutlin original data plus
retransmitted data
copy
free buffer space
R2
R2
l out
lin
Causescosts of congestion scenario 2
Host B
A
COMP 3825 - Kan Yang 6
6
Causescosts of congestion scenario 2
Idealization packets can be lost dropped at router due to full bufferssect sender only resends if packet known to be lost
lin original dataloutlin original data plus
retransmitted data
copy
no buffer spaceA
Host BCOMP 3825 - Kan Yang 7
7
Causescosts of congestion scenario 2
Idealization packets can be lost dropped at router due to full bufferssect sender only resends if packet known to be lost
lin original dataloutlin original data plus
retransmitted data
free buffer spaceA
Host BCOMP 3825 - Kan Yang 8
8
22620
3
Causescosts of congestion scenario 2
Realistic duplicatessect packets can be lost dropped at router
due to full bufferssect sender times out prematurely sending
two copies both of which are delivered
A
linloutlincopy
free buffer space
timeout
R2
R2lin
lout
when sending at R2 some packets are retransmissions including duplicated that are delivered
Host BCOMP 3825 - Kan Yang 9
9
Causescosts of congestion scenario 2Realistic duplicatessect packets can be lost dropped at router
due to full bufferssect sender times out prematurely sending
two copies both of which are delivered
R2
R2lin
lout
when sending at R2 some packets are retransmissions including duplicated that are delivered
ldquocostsrdquo of congestionsect more work (retrans) for given ldquogoodputrdquo (application-level throughput)sect unneeded retransmissions link carries multiple copies of pkt
bull decreasing goodput
COMP 3825 - Kan Yang 10
10
Causescosts of congestion scenario 3
sect four senderssect multihop pathssect timeoutretransmit
Q what happens as lin and linrsquo increase
finite shared output link buffers
Host A lout Host B
Host CHost D
lin original datalin original data plus
retransmitted data
A as red linrsquo increases all arriving blue pktsat upper queue are dropped blue throughput g 0
COMP 3825 - Kan Yang 11
11
another ldquocostrdquo of congestionsect when packet dropped any ldquoupstream transmission
capacityrdquo used for that packet was wasted
Causescosts of congestion scenario 3
C2
C2
l out
linrsquo
COMP 3825 - Kan Yang 12
12
22620
4
Learning Objectives
1 principles of congestion control
2 TCP congestion control
COMP 3825 - Kan Yang 13
13
TCP congestion control additive increase multiplicative decrease
sect approach sender increases transmission rate (window size) probing for usable bandwidth until loss occursbull additive increase increase cwnd (congestion window) by 1
MSS (maximum segment size) every RTT until loss detectedbull multiplicative decrease cut cwnd in half after loss
cwnd
TCP
send
er
cong
estio
n wi
ndow
size
AIMD saw toothbehavior probing
for bandwidth
additively increase window size helliphellip until loss occurs (then cut window in half)
timeCOMP 3825 - Kan Yang 14
14
TCP Congestion Control details
sect sender limits transmission
sect cwnd is dynamic function of perceived network congestion
TCP sending ratesect roughly send cwnd bytes wait
RTT for ACKS then send more bytes
last byteACKed sent not-
yet ACKed(ldquoin-flightrdquo)
last byte sent
cwnd
LastByteSent - LastByteAcked lt cwnd
sender sequence number space
rate ~~cwndRTT
bytessec
COMP 3825 - Kan Yang 15
15
TCP Slow Start sect when connection begins increase
rate exponentially until first loss eventbull initially cwnd = 1 MSSbull double cwnd every RTTbull done by incrementing cwnd for
every ACK received
sect summary initial rate is slow but ramps up exponentially fast
Host A
one segment
RTT
Host B
time
two segments
four segments
COMP 3825 - Kan Yang 16
16
22620
5
TCP detecting reacting to losssect loss indicated by timeout
bull cwnd set to 1 MSS bull window then grows exponentially (as in slow start) to threshold
then grows linearly
sect loss indicated by 3 duplicate ACKs TCP RENObull dup ACKs indicate network capable of delivering some segments bull cwnd is cut in half window then grows linearly
sect TCP Tahoe always sets cwnd to 1 (timeout or 3 duplicate acks)
COMP 3825 - Kan Yang 17
17
Fast RecoveryQ when should the exponential increase switch to linear A when cwnd gets to 12 of its value before timeout
Implementationsect variable ssthresh
(slow start threshold)sect on loss event ssthresh
is set to 12 of cwnd just before loss event
Check out the online interactive exercises for more examples httpgaiacsumassedukurose_rossinteractive
COMP 3825 - Kan Yang 18
18
TCP throughputbull avg TCP thruput as function of window size RTT
ndash ignore slow start assume always data to send
bull W window size (measured in bytes) where loss occursndash avg window size ( in-flight bytes) is frac34 Wndash avg thruput is 34W per RTT
W
W2
avg TCP thruput = 34
WRTT bytessec
COMP 3825 - Kan Yang 20
20
TCP Futures TCP over ldquolong fat pipesrdquo
bull example 1500 byte segments 100ms RTT want 10 Gbps throughputbull requires W = 83333 in-flight segmentsbull throughput in terms of segment loss probability L [Mathis 1997]
to achieve 10 Gbps throughput need a loss rate of L = 210-10 ndash a very small loss rate
bull new versions of TCP for high-speed
TCP throughput = 122 MSSRTT L
COMP 3825 - Kan Yang 21
21
22620
6
TCP Fairnessfairness goal if K TCP sessions share same bottleneck link of bandwidth R
each should have average rate of RK
TCP connection 1
bottleneckrouter
capacity RTCP connection 2
COMP 3825 - Kan Yang 22
22
Why is TCP fairtwo competing sessionssect additive increase gives slope of 1 as throughout increasessect multiplicative decrease decreases throughput proportionally
R
R
equal bandwidth share
Connection 1 throughput
Con
nect
ion
2 th
roug
hput
congestion avoidance additive increaseloss decrease window by factor of 2
congestion avoidance additive increaseloss decrease window by factor of 2
COMP 3825 - Kan Yang 23
23
Fairness (more)Fairness and UDPsect multimedia apps often do not use TCP
bull do not want rate throttled by congestion controlsect instead use UDP
bull send audiovideo at constant rate tolerate packet loss
Fairness parallel TCP connectionssect application can open multiple parallel connections between two hostssect web browsers do this sect eg link of rate R with 9 existing connections
bull new app asks for 1 TCP gets rate R10bull new app asks for 11 TCPs gets R2
COMP 3825 - Kan Yang 24
24
Explicit Congestion Notification (ECN)
network-assisted congestion controlsect two bits in IP header (ToS field) marked by network router to indicate congestionsect congestion indication carried to receiving hostsect receiver (seeing congestion indication in IP datagram) ) sets ECE (ECN-Echo) bit on
receiver-to-sender ACK segment to notify sender of congestion
source
applicationtransportnetwork
linkphysical
destination
applicationtransportnetwork
linkphysical
ECN=00 ECN=11
ECE=1
IP datagram
TCP ACK segment
COMP 3825 - Kan Yang 25
25
22620
3
Causescosts of congestion scenario 2
Realistic duplicatessect packets can be lost dropped at router
due to full bufferssect sender times out prematurely sending
two copies both of which are delivered
A
linloutlincopy
free buffer space
timeout
R2
R2lin
lout
when sending at R2 some packets are retransmissions including duplicated that are delivered
Host BCOMP 3825 - Kan Yang 9
9
Causescosts of congestion scenario 2Realistic duplicatessect packets can be lost dropped at router
due to full bufferssect sender times out prematurely sending
two copies both of which are delivered
R2
R2lin
lout
when sending at R2 some packets are retransmissions including duplicated that are delivered
ldquocostsrdquo of congestionsect more work (retrans) for given ldquogoodputrdquo (application-level throughput)sect unneeded retransmissions link carries multiple copies of pkt
bull decreasing goodput
COMP 3825 - Kan Yang 10
10
Causescosts of congestion scenario 3
sect four senderssect multihop pathssect timeoutretransmit
Q what happens as lin and linrsquo increase
finite shared output link buffers
Host A lout Host B
Host CHost D
lin original datalin original data plus
retransmitted data
A as red linrsquo increases all arriving blue pktsat upper queue are dropped blue throughput g 0
COMP 3825 - Kan Yang 11
11
another ldquocostrdquo of congestionsect when packet dropped any ldquoupstream transmission
capacityrdquo used for that packet was wasted
Causescosts of congestion scenario 3
C2
C2
l out
linrsquo
COMP 3825 - Kan Yang 12
12
22620
4
Learning Objectives
1 principles of congestion control
2 TCP congestion control
COMP 3825 - Kan Yang 13
13
TCP congestion control additive increase multiplicative decrease
sect approach sender increases transmission rate (window size) probing for usable bandwidth until loss occursbull additive increase increase cwnd (congestion window) by 1
MSS (maximum segment size) every RTT until loss detectedbull multiplicative decrease cut cwnd in half after loss
cwnd
TCP
send
er
cong
estio
n wi
ndow
size
AIMD saw toothbehavior probing
for bandwidth
additively increase window size helliphellip until loss occurs (then cut window in half)
timeCOMP 3825 - Kan Yang 14
14
TCP Congestion Control details
sect sender limits transmission
sect cwnd is dynamic function of perceived network congestion
TCP sending ratesect roughly send cwnd bytes wait
RTT for ACKS then send more bytes
last byteACKed sent not-
yet ACKed(ldquoin-flightrdquo)
last byte sent
cwnd
LastByteSent - LastByteAcked lt cwnd
sender sequence number space
rate ~~cwndRTT
bytessec
COMP 3825 - Kan Yang 15
15
TCP Slow Start sect when connection begins increase
rate exponentially until first loss eventbull initially cwnd = 1 MSSbull double cwnd every RTTbull done by incrementing cwnd for
every ACK received
sect summary initial rate is slow but ramps up exponentially fast
Host A
one segment
RTT
Host B
time
two segments
four segments
COMP 3825 - Kan Yang 16
16
22620
5
TCP detecting reacting to losssect loss indicated by timeout
bull cwnd set to 1 MSS bull window then grows exponentially (as in slow start) to threshold
then grows linearly
sect loss indicated by 3 duplicate ACKs TCP RENObull dup ACKs indicate network capable of delivering some segments bull cwnd is cut in half window then grows linearly
sect TCP Tahoe always sets cwnd to 1 (timeout or 3 duplicate acks)
COMP 3825 - Kan Yang 17
17
Fast RecoveryQ when should the exponential increase switch to linear A when cwnd gets to 12 of its value before timeout
Implementationsect variable ssthresh
(slow start threshold)sect on loss event ssthresh
is set to 12 of cwnd just before loss event
Check out the online interactive exercises for more examples httpgaiacsumassedukurose_rossinteractive
COMP 3825 - Kan Yang 18
18
TCP throughputbull avg TCP thruput as function of window size RTT
ndash ignore slow start assume always data to send
bull W window size (measured in bytes) where loss occursndash avg window size ( in-flight bytes) is frac34 Wndash avg thruput is 34W per RTT
W
W2
avg TCP thruput = 34
WRTT bytessec
COMP 3825 - Kan Yang 20
20
TCP Futures TCP over ldquolong fat pipesrdquo
bull example 1500 byte segments 100ms RTT want 10 Gbps throughputbull requires W = 83333 in-flight segmentsbull throughput in terms of segment loss probability L [Mathis 1997]
to achieve 10 Gbps throughput need a loss rate of L = 210-10 ndash a very small loss rate
bull new versions of TCP for high-speed
TCP throughput = 122 MSSRTT L
COMP 3825 - Kan Yang 21
21
22620
6
TCP Fairnessfairness goal if K TCP sessions share same bottleneck link of bandwidth R
each should have average rate of RK
TCP connection 1
bottleneckrouter
capacity RTCP connection 2
COMP 3825 - Kan Yang 22
22
Why is TCP fairtwo competing sessionssect additive increase gives slope of 1 as throughout increasessect multiplicative decrease decreases throughput proportionally
R
R
equal bandwidth share
Connection 1 throughput
Con
nect
ion
2 th
roug
hput
congestion avoidance additive increaseloss decrease window by factor of 2
congestion avoidance additive increaseloss decrease window by factor of 2
COMP 3825 - Kan Yang 23
23
Fairness (more)Fairness and UDPsect multimedia apps often do not use TCP
bull do not want rate throttled by congestion controlsect instead use UDP
bull send audiovideo at constant rate tolerate packet loss
Fairness parallel TCP connectionssect application can open multiple parallel connections between two hostssect web browsers do this sect eg link of rate R with 9 existing connections
bull new app asks for 1 TCP gets rate R10bull new app asks for 11 TCPs gets R2
COMP 3825 - Kan Yang 24
24
Explicit Congestion Notification (ECN)
network-assisted congestion controlsect two bits in IP header (ToS field) marked by network router to indicate congestionsect congestion indication carried to receiving hostsect receiver (seeing congestion indication in IP datagram) ) sets ECE (ECN-Echo) bit on
receiver-to-sender ACK segment to notify sender of congestion
source
applicationtransportnetwork
linkphysical
destination
applicationtransportnetwork
linkphysical
ECN=00 ECN=11
ECE=1
IP datagram
TCP ACK segment
COMP 3825 - Kan Yang 25
25
22620
4
Learning Objectives
1 principles of congestion control
2 TCP congestion control
COMP 3825 - Kan Yang 13
13
TCP congestion control additive increase multiplicative decrease
sect approach sender increases transmission rate (window size) probing for usable bandwidth until loss occursbull additive increase increase cwnd (congestion window) by 1
MSS (maximum segment size) every RTT until loss detectedbull multiplicative decrease cut cwnd in half after loss
cwnd
TCP
send
er
cong
estio
n wi
ndow
size
AIMD saw toothbehavior probing
for bandwidth
additively increase window size helliphellip until loss occurs (then cut window in half)
timeCOMP 3825 - Kan Yang 14
14
TCP Congestion Control details
sect sender limits transmission
sect cwnd is dynamic function of perceived network congestion
TCP sending ratesect roughly send cwnd bytes wait
RTT for ACKS then send more bytes
last byteACKed sent not-
yet ACKed(ldquoin-flightrdquo)
last byte sent
cwnd
LastByteSent - LastByteAcked lt cwnd
sender sequence number space
rate ~~cwndRTT
bytessec
COMP 3825 - Kan Yang 15
15
TCP Slow Start sect when connection begins increase
rate exponentially until first loss eventbull initially cwnd = 1 MSSbull double cwnd every RTTbull done by incrementing cwnd for
every ACK received
sect summary initial rate is slow but ramps up exponentially fast
Host A
one segment
RTT
Host B
time
two segments
four segments
COMP 3825 - Kan Yang 16
16
22620
5
TCP detecting reacting to losssect loss indicated by timeout
bull cwnd set to 1 MSS bull window then grows exponentially (as in slow start) to threshold
then grows linearly
sect loss indicated by 3 duplicate ACKs TCP RENObull dup ACKs indicate network capable of delivering some segments bull cwnd is cut in half window then grows linearly
sect TCP Tahoe always sets cwnd to 1 (timeout or 3 duplicate acks)
COMP 3825 - Kan Yang 17
17
Fast RecoveryQ when should the exponential increase switch to linear A when cwnd gets to 12 of its value before timeout
Implementationsect variable ssthresh
(slow start threshold)sect on loss event ssthresh
is set to 12 of cwnd just before loss event
Check out the online interactive exercises for more examples httpgaiacsumassedukurose_rossinteractive
COMP 3825 - Kan Yang 18
18
TCP throughputbull avg TCP thruput as function of window size RTT
ndash ignore slow start assume always data to send
bull W window size (measured in bytes) where loss occursndash avg window size ( in-flight bytes) is frac34 Wndash avg thruput is 34W per RTT
W
W2
avg TCP thruput = 34
WRTT bytessec
COMP 3825 - Kan Yang 20
20
TCP Futures TCP over ldquolong fat pipesrdquo
bull example 1500 byte segments 100ms RTT want 10 Gbps throughputbull requires W = 83333 in-flight segmentsbull throughput in terms of segment loss probability L [Mathis 1997]
to achieve 10 Gbps throughput need a loss rate of L = 210-10 ndash a very small loss rate
bull new versions of TCP for high-speed
TCP throughput = 122 MSSRTT L
COMP 3825 - Kan Yang 21
21
22620
6
TCP Fairnessfairness goal if K TCP sessions share same bottleneck link of bandwidth R
each should have average rate of RK
TCP connection 1
bottleneckrouter
capacity RTCP connection 2
COMP 3825 - Kan Yang 22
22
Why is TCP fairtwo competing sessionssect additive increase gives slope of 1 as throughout increasessect multiplicative decrease decreases throughput proportionally
R
R
equal bandwidth share
Connection 1 throughput
Con
nect
ion
2 th
roug
hput
congestion avoidance additive increaseloss decrease window by factor of 2
congestion avoidance additive increaseloss decrease window by factor of 2
COMP 3825 - Kan Yang 23
23
Fairness (more)Fairness and UDPsect multimedia apps often do not use TCP
bull do not want rate throttled by congestion controlsect instead use UDP
bull send audiovideo at constant rate tolerate packet loss
Fairness parallel TCP connectionssect application can open multiple parallel connections between two hostssect web browsers do this sect eg link of rate R with 9 existing connections
bull new app asks for 1 TCP gets rate R10bull new app asks for 11 TCPs gets R2
COMP 3825 - Kan Yang 24
24
Explicit Congestion Notification (ECN)
network-assisted congestion controlsect two bits in IP header (ToS field) marked by network router to indicate congestionsect congestion indication carried to receiving hostsect receiver (seeing congestion indication in IP datagram) ) sets ECE (ECN-Echo) bit on
receiver-to-sender ACK segment to notify sender of congestion
source
applicationtransportnetwork
linkphysical
destination
applicationtransportnetwork
linkphysical
ECN=00 ECN=11
ECE=1
IP datagram
TCP ACK segment
COMP 3825 - Kan Yang 25
25
22620
5
TCP detecting reacting to losssect loss indicated by timeout
bull cwnd set to 1 MSS bull window then grows exponentially (as in slow start) to threshold
then grows linearly
sect loss indicated by 3 duplicate ACKs TCP RENObull dup ACKs indicate network capable of delivering some segments bull cwnd is cut in half window then grows linearly
sect TCP Tahoe always sets cwnd to 1 (timeout or 3 duplicate acks)
COMP 3825 - Kan Yang 17
17
Fast RecoveryQ when should the exponential increase switch to linear A when cwnd gets to 12 of its value before timeout
Implementationsect variable ssthresh
(slow start threshold)sect on loss event ssthresh
is set to 12 of cwnd just before loss event
Check out the online interactive exercises for more examples httpgaiacsumassedukurose_rossinteractive
COMP 3825 - Kan Yang 18
18
TCP throughputbull avg TCP thruput as function of window size RTT
ndash ignore slow start assume always data to send
bull W window size (measured in bytes) where loss occursndash avg window size ( in-flight bytes) is frac34 Wndash avg thruput is 34W per RTT
W
W2
avg TCP thruput = 34
WRTT bytessec
COMP 3825 - Kan Yang 20
20
TCP Futures TCP over ldquolong fat pipesrdquo
bull example 1500 byte segments 100ms RTT want 10 Gbps throughputbull requires W = 83333 in-flight segmentsbull throughput in terms of segment loss probability L [Mathis 1997]
to achieve 10 Gbps throughput need a loss rate of L = 210-10 ndash a very small loss rate
bull new versions of TCP for high-speed
TCP throughput = 122 MSSRTT L
COMP 3825 - Kan Yang 21
21
22620
6
TCP Fairnessfairness goal if K TCP sessions share same bottleneck link of bandwidth R
each should have average rate of RK
TCP connection 1
bottleneckrouter
capacity RTCP connection 2
COMP 3825 - Kan Yang 22
22
Why is TCP fairtwo competing sessionssect additive increase gives slope of 1 as throughout increasessect multiplicative decrease decreases throughput proportionally
R
R
equal bandwidth share
Connection 1 throughput
Con
nect
ion
2 th
roug
hput
congestion avoidance additive increaseloss decrease window by factor of 2
congestion avoidance additive increaseloss decrease window by factor of 2
COMP 3825 - Kan Yang 23
23
Fairness (more)Fairness and UDPsect multimedia apps often do not use TCP
bull do not want rate throttled by congestion controlsect instead use UDP
bull send audiovideo at constant rate tolerate packet loss
Fairness parallel TCP connectionssect application can open multiple parallel connections between two hostssect web browsers do this sect eg link of rate R with 9 existing connections
bull new app asks for 1 TCP gets rate R10bull new app asks for 11 TCPs gets R2
COMP 3825 - Kan Yang 24
24
Explicit Congestion Notification (ECN)
network-assisted congestion controlsect two bits in IP header (ToS field) marked by network router to indicate congestionsect congestion indication carried to receiving hostsect receiver (seeing congestion indication in IP datagram) ) sets ECE (ECN-Echo) bit on
receiver-to-sender ACK segment to notify sender of congestion
source
applicationtransportnetwork
linkphysical
destination
applicationtransportnetwork
linkphysical
ECN=00 ECN=11
ECE=1
IP datagram
TCP ACK segment
COMP 3825 - Kan Yang 25
25
22620
6
TCP Fairnessfairness goal if K TCP sessions share same bottleneck link of bandwidth R
each should have average rate of RK
TCP connection 1
bottleneckrouter
capacity RTCP connection 2
COMP 3825 - Kan Yang 22
22
Why is TCP fairtwo competing sessionssect additive increase gives slope of 1 as throughout increasessect multiplicative decrease decreases throughput proportionally
R
R
equal bandwidth share
Connection 1 throughput
Con
nect
ion
2 th
roug
hput
congestion avoidance additive increaseloss decrease window by factor of 2
congestion avoidance additive increaseloss decrease window by factor of 2
COMP 3825 - Kan Yang 23
23
Fairness (more)Fairness and UDPsect multimedia apps often do not use TCP
bull do not want rate throttled by congestion controlsect instead use UDP
bull send audiovideo at constant rate tolerate packet loss
Fairness parallel TCP connectionssect application can open multiple parallel connections between two hostssect web browsers do this sect eg link of rate R with 9 existing connections
bull new app asks for 1 TCP gets rate R10bull new app asks for 11 TCPs gets R2
COMP 3825 - Kan Yang 24
24
Explicit Congestion Notification (ECN)
network-assisted congestion controlsect two bits in IP header (ToS field) marked by network router to indicate congestionsect congestion indication carried to receiving hostsect receiver (seeing congestion indication in IP datagram) ) sets ECE (ECN-Echo) bit on
receiver-to-sender ACK segment to notify sender of congestion
source
applicationtransportnetwork
linkphysical
destination
applicationtransportnetwork
linkphysical
ECN=00 ECN=11
ECE=1
IP datagram
TCP ACK segment
COMP 3825 - Kan Yang 25
25
