Transcript CMPT 880: Internet Architectures and Protocols

School of Computing Science
Simon Fraser University
CMPT 771/471: Internet Architecture and Protocols
Transport Layer
Instructor: Dr. Mohamed Hefeeda
1
Review of Basic Networking Concepts
 Internet structure
 Protocol layering and encapsulation
 Internet services and socket programming
 Network Layer
 Network types: Circuit switching, Packet switching
 Addressing, Forwarding, Routing
 Transport layer
 Reliability, congestion and flow control
 TCP, UDP
 Link Layer
 Multiple Access Protocols
 Ethernet
2
Transport services and protocols
 provide logical communication
between app processes
running on different hosts
 transport protocols run in end
systems
 send side: breaks app
messages into segments,
passes to network layer
 rcv side: reassembles
segments into messages,
passes to app layer
 more than one transport
protocol available to apps
 Internet: TCP and UDP
application
transport
network
data link
physical
network
data link
physical
network
data link
physical
network
data link
physical
network
data link
physical
network
data link
physical
application
transport
network
data link
physical
3
Transport vs. network layer
 network layer: logical
communication between
hosts
Household analogy:
 transport layer: logical
communication between
processes
 processes = kids
 relies on, enhances,
network layer services
12 kids sending letters to 12
kids
 app messages = letters in
envelopes
 hosts = houses
 transport protocol = Ann
and Bill
 network-layer protocol =
postal service
4
Multiplexing/demultiplexing
Multiplexing at send host:
gathering data from multiple
sockets, enveloping data with
header (later used for
demultiplexing)
Demultiplexing at rcv host:
delivering received segments
to correct socket
= socket
application
transport
network
link
= process
P3
P1
P1
application
transport
network
P2
P4
application
transport
network
link
link
physical
host 1
physical
host 2
physical
host 3
5
Connectionless demux
P2
client
IP: A
P1
P1
P3
SP: 9157
DP: 6428
SP: 6428
SP: 6428
DP: 9157
DP: 5775
SP: 5775
server
IP: C
DP: 6428
Client
IP:B
UDP socket identified by: (dst IP, dst Port) 
datagrams with different src IPs and/or src ports are directed to same socket
6
Connection-oriented demux (cont)
P1
P4
P5
P2
P6
P1P3
SP: 5775
DP: 80
S-IP: B
D-IP:C
SP: 9157
client
IP: A
DP: 80
S-IP: A
D-IP:C
SP: 9157
server
IP: C
DP: 80
S-IP: B
D-IP:C
Client
IP:B
 TCP socket identified by 4-tuple: (src IP, src Port, dst IP, dst Port)
7
UDP: User Datagram Protocol [RFC 768]
 “no frills,” “bare bones”
Internet transport protocol
 “best effort” service, UDP
segments may be:
 lost
 delivered out of order to
app
 Connectionless:
 no handshaking
between UDP sender,
receiver
 each UDP segment
handled independently
of others
Why is there a UDP?
 no connection establishment
(which can add delay)
 simple: no connection state
at sender, receiver
 small segment header
 no congestion control: UDP
can blast away as fast as
desired
8
UDP
 often used for streaming
multimedia apps
 loss tolerant
Length, in
 rate sensitive
bytes of UDP
 other UDP uses
 DNS
 SNMP
 reliable transfer over UDP:
add reliability at application
layer
 application-specific
error recovery!
segment,
including
header
32 bits
source port #
dest port #
length
checksum
Application
data
(message)
UDP segment format
9
Reliable data transfer
 important in application, transport, and link layers
 top-10 list of important networking topics!
 characteristics of unreliable channel will determine complexity of
reliable data transfer protocol (rdt)
10
Pipelined (Sliding Window) Protocols
Pipelining: sender allows multiple, “in-flight”, yet-to-beacknowledged pkts
 range of sequence numbers must be increased
 buffering at sender and/or receiver
 Two generic forms of pipelined protocols: go-Back-N, selective
repeat
11
Go-Back-N
Sender:
 k-bit seq # in pkt header
 “window” of up to N, consecutive unack’ed pkts allowed
 ACK(n): ACKs all pkts up to, including seq # n -- cumulative ACK
 may receive duplicate ACKs (see receiver)
 timer for each in-flight pkt
 timeout(n): retransmit pkt n and all higher seq # pkts in window
 i.e., go back to n
12
GBN in
action
Go back to 2
Window size, N = 4
13
Go-Back-N
 Do you see potential problems with GBN?
 Consider high-speed links with long delays
 (called large bandwidth-delay product pipes)
 GBN can fill that pipe by having large N 
 many unACKed pkts could be in the pipe
 A single lost pkt could cause a re-transmission of a huge
number (up to N) of pkts  waste of bandwidth
 Solutions??
14
Selective Repeat
 receiver individually acknowledges all correctly received
pkts
 buffers pkts, as needed, for eventual in-order delivery to
upper layer
 sender only resends pkts for which ACK not received
 sender timer for each unACKed pkt
 sender window
 N consecutive seq #’s
 again limits seq #s of sent, unACKed pkts
15
Selective repeat: sender, receiver windows
16
TCP: Overview
RFCs: 793, 1122, 1323, 2018, 2581
 point-to-point:
 full duplex data:
 one sender, one receiver
 bi-directional data flow in
same connection
 MSS: maximum segment
size
 reliable, in-order byte
stream:
 no “message boundaries”
 pipelined:
 connection-oriented:
 handshaking (exchange
of control msgs) init’s
sender, receiver state
before data exchange
 TCP congestion and flow
control set window size
 send & receive buffers
 flow controlled:
socket
door
application
writes data
application
reads data
TCP
send buffer
TCP
receive buffer
 sender will not
overwhelm receiver
socket
door
segment
17
TCP segment structure
32 bits
URG: urgent data
(generally not used)
ACK: ACK #
valid
PSH: push data now
(generally not used)
RST, SYN, FIN:
connection estab
(setup, teardown
commands)
Internet
checksum
(as in UDP)
source port #
dest port #
sequence number
acknowledgement number
head not
UA P R S F
len used
checksum
Receive window
Urg data pnter
Options (variable length)
counting
by bytes
of data
(not segments!)
# bytes
rcvr willing
to accept
application
data
(variable length)
18
TCP reliable data transfer
 TCP creates rdt service
on top of IP’s unreliable
service
 Pipelined segments
 Cumulative acks
 TCP uses single
retransmission timer
 Retransmissions are
triggered by:
 timeout events
 duplicate acks
 Initially consider
simplified TCP sender:
 ignore duplicate acks
 ignore flow control,
congestion control
19
TCP sender events:
data rcvd from app:
timeout:
 Create segment with seq #
 retransmit segment that
caused timeout
 seq # is byte-stream
number of first data byte
in segment
 start timer if not already
running (think of timer as
for oldest unacked
segment)
 expiration interval:
TimeOutInterval
 restart timer
Ack rcvd:
 If acknowledges
previously unacked
segments
 update what is known to
be acked
 start timer if there are
outstanding segments
20
NextSeqNum = InitialSeqNum
SendBase = InitialSeqNum
loop (forever) {
switch(event)
event: data received from application above
create TCP segment with sequence number NextSeqNum
if (timer currently not running)
start timer
pass segment to IP
NextSeqNum = NextSeqNum + length(data)
TCP
sender
(simplified)
event: timer timeout
retransmit not-yet-acknowledged segment with
smallest sequence number
start timer
event: ACK received, with ACK field value of y
if (y > SendBase) {
SendBase = y
if (there are currently not-yet-acknowledged segments)
start timer
}
} /* end of loop forever */
21
TCP: retransmission scenarios
Host A
X
loss
Sendbase
= 100
SendBase
= 120
SendBase
= 100
time
SendBase
= 120
lost ACK scenario
Host B
Seq=92 timeout
Host B
Seq=92 timeout
timeout
Host A
time
premature timeout
22
TCP retransmission scenarios (more)
timeout
Host A
Host B
X
loss
SendBase
= 120
time
Cumulative ACK scenario
23
TCP Round Trip Time and Timeout
 If TCP timeout is
 too short: premature timeout  unnecessary
retransmissions
 too long: slow reaction to segment loss
Q: how to set TCP timeout value?
 Based on Round Trip Time (RTT), but RTT itself varies with
time!
 We need to estimate current RTT
 RTT Estimation
 SampleRTT: measured time from segment transmission
until ACK receipt
 ignore retransmissions
 SampleRTT will vary, want estimated RTT “smoother”
 average several recent measurements, not just current
SampleRTT
24
TCP Round Trip Time and Timeout
EstimatedRTT = (1- )*EstimatedRTT + *SampleRTT
 Exponential weighted moving average
 influence of past sample decreases exponentially fast
 typical value:  = 0.125
25
Example RTT estimation:
RTT: gaia.cs.umass.edu to fantasia.eurecom.fr
350
RTT (milliseconds)
300
250
200
150
100
1
8
15
22
29
36
43
50
57
64
71
78
85
92
99
106
time (seconnds)
SampleRTT
Estimated RTT
26
TCP Round Trip Time and Timeout
Setting the timeout
 EstimtedRTT plus safety margin
 large variation in EstimatedRTT -> larger safety margin
 first estimate how much SampleRTT deviates from
EstimatedRTT:
DevRTT = (1-)*DevRTT +
*|SampleRTT - EstimatedRTT|
(typically,  = 0.25)
Then set timeout interval:
TimeoutInterval = EstimatedRTT + 4*DevRTT
27
Fast Retransmit
 Time-out period often
relatively long:
 long delay before
resending lost packet
 Detect lost segments via
duplicate ACKs.
 Sender often sends
many segments back-toback
 If segment is lost, there
will likely be many
duplicate ACKs.
 If sender receives 3 ACKs
for the same data, it
supposes that segment
after ACKed data was lost:
 fast retransmit: resend
segment before timer
expires
28
TCP Connection Management: opening
 TCP: 3-way handshake
Step 1: client host sends TCP SYN segment
client
to server
 specifies initial seq #
conn.
 no data
request
server
conn.
granted
Step 2: server host receives SYN, replies
with SYNACK segment
 server allocates buffers
 specifies server initial seq. #
Step 3: client receives SYNACK, replies
with ACK segment, which may contain
data
A. SYN Flood
DoS attack
Q. How would a hacker exploit TCP 3-way handshake to bring a server down?
29
TCP Connection Management: closing
Step 1: client end system sends
TCP FIN segment to server
client
server
closing
Step 2: server receives FIN, replies
with ACK. Closes connection,
sends FIN
 Enters “timed wait” – may
need to re-send ACK to
received FINs
timed wait
Step 3: client receives FIN, replies
with ACK
closing
closed
closed
Step 4: server, receives ACK
Connection closed
30
TCP Connection Management
TCP server
lifecycle
TCP client
lifecycle
31
TCP Flow Control
 receive side of TCP
connection has a receive
buffer:
flow control
sender won’t overflow
receiver’s buffer by
transmitting too much,
too fast
 speed-matching service:
matching the send rate to
the receiving app’s drain
rate
 app process may be slow
at reading from buffer
32
TCP Flow control: how it works
 Rcvr advertises spare
room by including value of
RcvWindow in segments
 Sender limits unACKed
data to RcvWindow
(Suppose TCP receiver discards
out-of-order segments)
 spare room in buffer
 guarantees receive
buffer doesn’t overflow
= RcvWindow
= RcvBuffer-[LastByteRcvd LastByteRead]
33
Congestion Control
 Congestion: sources send too much data for network to
handle
 different from flow control, which is e2e
 Congestion results in …
 lost packets (buffer overflow at routers)
• more work (retransmissions) for given “goodput”
 long delays (queueing in router buffers)
• Premature (unneeded) retransmissions
 Waste of upstream links’ capacity
• Pkt traversed several links, then dropped at
congested router
34
Approaches towards congestion control
Two broad approaches towards congestion control:
End-end congestion control:
 no explicit feedback from
network
 congestion inferred from
end-system observed loss,
delay
 approach taken by TCP
Network-assisted congestion
control:
 routers provide feedback to
end systems
 single bit indicating
congestion (SNA,
DECbit, TCP/IP ECN,
ATM)
 explicit rate sender
should send at
35
TCP congestion control: Approach
 Approach: probe for usable bandwidth in network
 increase transmission rate until loss occurs then decrease
 Additive increase, multiplicative decrease (AIMD)
congestion
window
Saw tooth
behavior: probing
for bandwidth
Rate (CongWin)
24 Kbytes
16 Kbytes
8 Kbytes
time
time
36
TCP Congestion Control
 Sender keeps a new variable, Congestion Window (CongWin),
and limits unacked bytes to:
LastByteSent - LastByteAcked  min {CongWin, RcvWin}
 For our discussion: assume RcvWin is large enough
 Roughly, what is the sending rate as a function of CongWin?
 Ignore loss and transmission delay
 Rate = CongWin/RTT
(bytes/sec)
 So, rate and CongWin are somewhat synonymous
37
TCP Congestion Control
 Congestion occurs at routers (inside the network)
 Routers do not provide any feedback to TCP
 How can TCP infer congestion?
 From its symptoms: timeout or duplicate acks
 Define loss event ≡ timeout or 3 duplicate acks
 TCP decreases its CongWin (rate) after a loss event
 TCP Congestion Control Algorithm: three components
 AIMD: additive increase, multiplicative decrease
 slow start
 Reaction to timeout events
38
AIMD
 additive increase: (congestion avoidance phase)
 increase CongWin by 1 MSS every RTT until loss detected
 TCP increases CongWin by: MSS x (MSS/CongWin) for every
ACK received
 Ex. MSS = 1,460 bytes and CongWin = 14,600 bytes
 With every ACK, CongWin is increased by 146 bytes
 multiplicative decrease:
 cut CongWin in half after loss
congestion
window
CongWin
24 Kbytes
16 Kbytes
8 Kbytes
time
39
TCP Slow Start
 When connection begins, CongWin = 1 MSS
 Example: MSS = 500 bytes & RTT = 200 msec
 initial rate = CongWin/RTT = 20 kbps
 available bandwidth may be >> MSS/RTT
 desirable to quickly ramp up to respectable rate
 Slow start:
 When connection begins, increase rate exponentially fast until
first loss event. How can we do that?
 double CongWin every RTT. How?
 Increment CongWin by 1 MSS for every ACK received
40
TCP Slow Start (cont’d)
 Increment CongWin by 1
MSS for every ACK
Host B
RTT
Host A
 Summary: initial rate is
slow but ramps up
exponentially fast
time
41
Reaction to a Loss event
 TCP Tahoe (Old)
 Threshold = CongWin / 2
 Set CongWin = 1
 Slow start till threshold
 Then Additive Increase
// congestion avoidance
 TCP Reno (most current TCP implementations)
 If 3 dup acks
// fast retransmit
• Threshold = CongWin / 2
• Set CongWin = Threshold // fast recovery
• Additive Increase
 Else
// timeout
• Same as TCP Tahoe
42
Reaction to a Loss event (cont’d)
3 dup acks
 Why differentiate between 3 dup acks and timeout?
 3 dup ACKs indicate network capable of
delivering some segments
 timeout indicates a “more alarming” congestion scenario
43
TCP Congestion Control: Summary
 Initially
 Threshold is set to large value (65 Kbytes), has no effect
 CongWin = 1 MSS
 Slow Start (SS): CongWin grows exponentially
 till a loss event occurs (timeout or 3 dup ack) or reaches
Threshold
 Congestion Avoidance (CA): CongWin grows linearly
 3 duplicate ACK occurs:
 Threshold = CongWin/2; CongWin = Threshold; CA
 Timeout occurs:
 Threshold = CongWin/2; CongWin = 1 MSS; SS till
Threshold
44
TCP Throughput Analysis
 Understand the fundamental relationship between
 Packet loss probability,
 RTT, and
 TCP performance (throughput)
 We present simple model, with several assumptions
 Yet it still provides useful insights
 See Ch 5 of [HJ04] for a summary of more detailed models
with references to the original papers
45
TCP Throughput Analysis
 Any TCP model must capture
 Window Dynamics (internal and deterministic)
• Controlled internally by the TCP algorithms.
• Depends on the particular flavor of TCP
• We assume TCP Reno (the most common)
 Packet Loss Process (external and uncertain)
• Models the aggregate of network conditions at all
nodes in the TCP connection path
• Typically modeled as a Stochastic Process with
probability p that a packet loss occurs
• TCP responds by reducing the window size
 We usually analyze the steady state
 Ignore the slow start phase (transient)
 Although many connections finish within slow start,
because they send only a few kilobytes
46
Notations
 X(t): Throughput at time t (transmission rate)
 W(t): window size at time t
 RTT: Round Trip Time
X(t) = W(t)/RTT
 What does the above equation implicitly assume?
 Increasing X(t) has negligible effects on the queuing
delay in the network  RTT remains constant
47
Simple (Periodic) Model
loss occurs
 Packet losses occur with
constant probability p
W
 TCP window starts at W/2
grows to W, then halves,
repeat forever …
W/2
period
time (RTT)
 W(t) packets transmitted
each RTT
 W(t+1) = W(t) + 1 each
round until a loss occurs
48
Simple (Periodic) Model
T
Compute the steady state throughput as a function of
average loss probability p.
Average# of PacketsSent During a Period 1 / p
X ( p) 

Period Length
T
49
Simple (Periodic) Model
T
 T: period between detecting packet losses  T = RTT * W /2
 Now, we find W as a function of p. How?
 Compute the number of packets sent during a period and
equate it to 1/p. (Size of the green area):
W/2 * (W/2 + W) / 2 = 1/p  W = sqrt(8/3p)
50
Simple (Periodic) Model
 Inverse Square-Root-p Law
1
X ( p) 
RTT
3
2p
 TCP throughput is inversely proportional to
 RTT and square root of packet loss probability p
51
In More Realistic Models …
 Packet loss probability is not constant and is bursty
 Consider effect of duplicate ACKs and Timeouts
 Consider receiver window limit
52