MIT OpenCourseWare
http://ocw.mit.edu
6.033 Computer System Engineering
Spring 2009
For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.
L12: end to end layer
Dina Katabi
Some slides are from lectures by
Nick Mckeown, Ion Stoica, Frans
Kaashoek, Hari Balakrishnan, Sam
Madden, and Robert Morris
End-to-end layer
client
server
presentation Layer
stub
stub
End-to-end layer
RPC
D
Data
Header
Data
RPC
Header
Network Layer
H
D
H
D
D
H
H
D
D
H
Link Layer
H
Network layer provides
best effort service
• Packets may be:
•
•
•
•
•
Lossed
Delayed (jitter)
Duplicated
Reordered
…
• Problem: Inconvenient service for applications
• Solution: Design protocols for E2E modules
• Many protocols/modules possible, depending on requirements
This lecture: some E2E properties
• At most once
• At least once
• Exactly once?
• Sliding window
• Case study: TCP
• Tomorrow: Network File System (NFS)
At Least Once
client
server
DDaata
ta
an RTT
client
server
DDaata
ta
AACCKK
X
Timeout and
Retransmission
DDaata
ta
• Sender persistently sends until it receives an ack
• Challenges:
• Duplicate ACKs
• What value for timer
Duplicate ACK problem
Client
Server
RReeq 1
q1
timeout
AACCKK
RReeq 1
q1
RReeq 2
q2
AACCKK
RReeq 3
q3
• Problem: Request 2 is not delivered
• violates at-least once delivery
Solution: nonce
Client
N1
timeout
N2
Server
RReeq N
q N11
11
N
N
K
AACCK
RReeq N
q N11
RReeq N
q N22
NN11
K
AACCK
RReeq N
q N22
• Label request and ack with unique identifier that is never re-used
Engineering a nonce
Client
• Use sequence numbers
• Challenges:
• Wrap around?
• Failures?
1
timeout
2
Server
RReeq 1
q1
11
K
AACCK
RReeq 1
q1
RReeq 2
q2
11
K
AACCK
RReeq 2
q2
Timer value
• Fixed is bad. RTT changes depending on
congestion
• Pick a value that’s too big, wait too long to
retransmit a packet
• Pick a value too small, generates a duplicate
(retransmitted packet).
• Adapt the estimate of RTT Æ adaptive timeout
Adaptive Timeout:
Exponential weighted moving averages
• Samples S1, S2, S3, ..
• Algorithm
• EstimatedRTT = T0
• EstimatedRTT = α S + (1- α) EstimatedRTT
• where 0 ≤ α ≤ 1
• What values should one pick for α and T0?
• Adaptive timeout is also hard
At Most Once Challenges
client
1
2
server
RReeq 1
q1
OOkk 1
KK 1
C
AAC
rreeq 1
q1
11
K
AACCK
Process request 1
Process request 1
• Server shouldn’t process req 1
• Server should send result preferably
Idea: remember sequence number
client
1
2
server
RReeq 1
q1
OOkk 1
KK 1
C
AAC
rreeq 1
q1
Process request 1
Ok
1
Ok
OOkk 1
KK 1
C
AAC
Resend ACK 1
ACK
ACK11
• Server remembers also last few responses
Problem: failures
client
1
2
server
RReeq 1
q1
OOkk 1
KK 1
C
AAC
rreeq 1
q1
OOkk 1
KK 1
C
AAC
0
1
Ok
Ok
ACK
ACK11
0
1
Ok
Ok
ACK
ACK11
• Performed request 1 twice!
• How to maintain the last nonce per sender (tombstone)?
• Write to non-volatile storage?
• Move the problem? (e.g., different port number)
• Make probability of mistake small?
• How about exactly once? (Need transactions)
How fast should the sender sends?
Host A
Host B
DDaata
ta11
AACCKK
DDaata
ta22
• Waiting for acks is too slow
• Throughput is one
packet/RTT
• Say packet is 500 bytes
• RTT 100ms
• Æ Throughput = 40Kb/s,
Awful!
• Overlap pkt transmission
Send a window of packets
Host A
Host B
Send?
• Maybe because the
receiver is a slow
machine
ts
3 pk
5-7
2-4
O K,
• Assume the receiver
is the bottleneck
Idle
• Receiver needs to
tell the sender when
and how much it can
send
• The window
advances once all
previous packets are
acked Æ too slow
Sliding Window
Host A
Host B
• Senders advances
the window
whenever it
receives an ack Æ
sliding window
Send?
ts
3-5
2-4
O
pk
3
,
K
Idle
• But what is the
right value for the
window?
The Right Window Size
• Assume server is bottleneck
•
•
•
•
Goal: make idle time on server zero
Assume: server rate is B bytes/s
Window size = B x RTT
Danger: sequence number wrap around
• What if network is bottleneck?
• Many senders?
• Sharing?
• Next lecture
“Negative” ACK
Host A
D1
D2
D3
D4
Host B
X
k- 2
c
a
N
D2
D1
D3
• Minimize reliance on timer
• Add sequence numbers to
packets
• Send a Nack when the
receiver finds a hole in
the sequence numbers
• Difficulties
• Reordering
• Cannot eliminate acks,
because we need to ack the
last packet
E2E layer in Internet
HTTP, RTP, Sun RPC, …
Application
End-to-End
Layer
TCP or UDP
IP
Ethernet, WiFI, ...
Transport
Network
Link
The 4-layer Internet model
UDP
Transmission Control Protocol (TCP)
• Connection-oriented
• Delivers bytes atmost-once
• Bidirectional
Host A
x,?
Syn x
ac
• ACKs are piggybacked
x+1, y+1
Host B
yn y
s
,
1
k x+
D a ta x
+1, ac
k y+1
y, x+1
TCP header
Closing a TCP connection
Host A
x,y
Host B
fin x
y, x
+1
x
k
ac
fin y
timed wait
timeout
closed
ack y+
1
closed
(Start)
Connect/SYN (step 1 of the 3 way handshake)
Closed
Listen
SYN/SYN+ACK (step 2 of the 3 way handshake)
Close
Close
Listen
SYN
Received
SEND/SYN
RST
ACK
Data exchange occur
Established
Close/FIN
Close/FIN
SYN+ACK/ACK
(step 3 of the 3 way handshake)
FIN/ACK
Active close
FIN WAIT 1
SYN Sent
SYN/SYN+ACK (simultaneous open)
FIN/ACK
Passive close
Closing
Close WAIT
FIN+ACK/ACK
ACK
Close/FIN
ACK
FIN WAIT 2
FIN/ACK
(Go back to start)
Unusual event
TIMED WAIT
Last ACK
Timeout
Closed
Client/reciever path
ACK
Sender/server path
Figure by MIT OpenCourseWare.