Eifel Detection Algorithm*

1
Eifel Detection Algorithm

• Presented by
• Maitreya Natu
• CISC 856, Fall 2004
• natu_at_cis.udel.edu

Reiner Ludwig, M. Meyer, Ericsson Research,
April 2003, Request for Comments 3522
2
What is Eifel algorithm?

• Enhancement to TCPs error recovery scheme
• Eliminates retransmission ambiguity, there by
  solving problems caused by spurious timeouts and
  spurious retransmits
• Can improve end-to-end performance by several
  tens of percent
• Makes TCP truly wireless compatible

3
Timeout based retransmission
cwnd 8
ssthresh 16
ssthresh 8/2 4
Timeout
cwnd 1
Slow start
cwnd 2
. . .
cwnd 4

cwnd 16
Congestion avoidance

cwnd 17
4
Fast retransmit
ssthresh 16
A1
cwnd 8
D1

D8
A1
A1
3 Dup-ACKs
A1
D1
Swnd 437
ssthresh 8/2 4
cwnd 4
Fast retransmit fast recovery
A9
Di
cwnd 4
Dj

Congestion avoidance
Ai1
Aj1
cwnd 5
5
Spurious Timeout
ssthresh 16
D1
cwnd 8
D8
Timeout
6
Spurious Timeout
ssthresh 16
D1
cwnd 8
D8
ssthresh 8/2 4
D1
Timeout
cwnd 1
Slow start
7
Spurious Timeout
ssthresh 16
D1
cwnd 8
D8
A2
ssthresh 8/2 4
D1
Timeout
cwnd 1
Slow start
8
Spurious Timeout

• Q1 When does a spurious timeout occur?

ssthresh 16
D1
cwnd 8
D8
A2
ssthresh 8/2 4
D1
Timeout
cwnd 1
Slow start
Spurious timeout spurious retransmission
9
Spurious Timeout

• Q1 When does a spurious timeout occur?
• When RTT suddenly increases, to the extent that
  it exceeds the retransmission time out value that
  had been determined a prior
• Q2 What can cause RTT to increase suddenly?

ssthresh 16
D1
cwnd 8
D8
A2
ssthresh 8/2 4
D1
Timeout
cwnd 1
Slow start
10
Spurious Timeout

• Q1 When does a spurious timeout occur?
• When RTT suddenly increases, to the extent that
  it exceeds the retransmission time out value that
  had been determined a prior
• Q2 What can cause RTT to increase suddenly?
• Route changes
• Rapid increase in congestion

ssthresh 16
D1
cwnd 8
D8
A2
ssthresh 8/2 4
D1
Timeout
cwnd 1
Slow start
11
Spurious Timeout
ssthresh 16
D1
cwnd 8
D8
A2
ssthresh 8/2 4
D1
Timeout
cwnd 1
Slow start
12
Spurious Timeout
ssthresh 16
D1
cwnd 8
D8
A2
ssthresh 8/2 4
D1
Timeout
cwnd 1
D2
cwnd 2
D3
Slow start
13
Spurious Timeout
ssthresh 16
D1
cwnd 8
D8
A2
A3
A4
ssthresh 8/2 4
D1
Timeout
cwnd 1
D2
cwnd 2
D3
Slow start
14
Spurious Timeout

• Q3 How does it affect TCP performance
• Sender unnecessarily reduces the load
• Sender if forced into a go-back-N retransmission
  mode
• Can lead to real packet losses due to congestion
  caused by aggressive sender behavior

ssthresh 16
D1
cwnd 8
D8
A2
A3
A4
ssthresh 8/2 4
D1
Timeout
cwnd 1
D2
cwnd 2
D3
cwnd 4
Slow start
15
Spurious fast retransmit
ssthresh 16
A1
D1
cwnd 8
D8

A1
A1
A1
3 Dup-ACKs
16
Spurious fast retransmit
ssthresh 16
A1
D1
cwnd 8
D8

A1
A1
A1
3 Dup-ACKs
ssthresh 4
D1
cwnd 4 swnd 7
Fast retransmit fast recovery
17
Spurious fast retransmit
ssthresh 16
A1
D1
cwnd 8
D8

A1
A1
A1
3 Dup-ACKs
ssthresh 4
D1
cwnd 4 swnd 7
A9
Fast retransmit fast recovery
18
Spurious fast retransmit
ssthresh 16
A1
D1
cwnd 8
D8

A1
A1
A1
3 Dup-ACKs
ssthresh 4
D1
cwnd 4 swnd 7
A9
Fast retransmit fast recovery
Spurious fast retransmit
19
Spurious fast retransmit
ssthresh 16
A1

• Q1 When does a spurious fast retransmit occur?
• Occur due to reordering of packets beyond the
  DUP-ACK threshold
• Frequency of occurrence depends on path
  properties
• E.g. due to load-balancing on a routers
  inter-connected via multiple links

D1
cwnd 8
D8

A1
A1
A1
3 Dup-ACKs
ssthresh 4
D1
cwnd 4 swnd 7
A9
Fast retransmit fast recovery
Spurious fast retransmit
20
Spurious fast retransmit
ssthresh 16
A1
D1
cwnd 8
D8

A1
A1
A1
3 Dup-ACKs
ssthresh 4
D1
cwnd 4 swnd 7
A9
Fast retransmit fast recovery
21
Spurious fast retransmit
ssthresh 16
A1
D1
cwnd 8
D8

A1
A1
A1
3 Dup-ACKs
ssthresh 4
D1
cwnd 4 swnd 7
Fast retransmit fast recovery
A9
cwnd 4


Congestion avoidance
cwnd 5
22
Spurious fast retransmit
ssthresh 16
A1

• Q2 How does it affect TCP performance
• Causes the sender to unnecessarily reduce its
  load
• Causes the sender to unnecessarily retransmit

D1
cwnd 8
D8

A1
A1
A1
3 Dup-ACKs
ssthresh 4
D1
cwnd 4 swnd 7
Fast retransmit fast recovery
A9
cwnd 4


Congestion avoidance
cwnd 5
23
Spurious Timeout

• Problem
• Senders inability to distinguish an ACK for
  original transmission from that of retransmission

ssthresh 16
D1
cwnd 8
D8
A2
ssthresh 8/2 4
D1
Timeout
cwnd 1
Slow start
Spurious timeout spurious retransmission
24
Spurious fast retransmit
ssthresh 16
A1

• Problem
• Senders inability to distinguish an ACK for
  original transmission from that of retransmission

D1
cwnd 8
D8

A1
A1
A1
3 Dup-ACKs
ssthresh 4
D1
cwnd 4 swnd 7
A9
Fast retransmit fast recovery
Spurious fast retransmit
25
Problem Solution

• Problem
• Retransmission ambiguity
• Senders inability to distinguish an ACK for
  original transmission from the ACK for
  retransmission
• Solution
• Eliminate retransmission ambiguity
• Use extra information in the ACKs to distinguish
  an ACK for original transmission from an ACK for
  retransmission

26
Review of TCP timestamp optionRFC 1323
Time at receiver
Time at sender
100
100
200
200, 100
300
300, 200
301
301, 200
400
Timestamp of the most recent data segment that
advanced the window is echoed (here 301)
400, 301
4.4BSD increments the timestamp clock once every
500ms and this timestamp clock is reset to 0 on a
reboot - TCP/IP ILLUS v1, p349
27
Eliminating retransmission ambiguity
A , 500, xxx
500
100
D4, 100, 500

D11, 100, 500
Time-out
ts_first_rexmit 400
400
D4, 400, 500
28
Eliminating retransmission ambiguity
A , 500, xxx
500
100
D4, 100, 500

D11, 100, 500
Time-out
501
A5, 501, 100
ts_first_rexmit 400
400
D4, 400, 500
29
Eliminating retransmission ambiguity
A , 500, xxx
500
100
D4, 100, 500

D11, 100, 500
Time-out
501
A5, 501, 100
ts_first_rexmit 400
400
D4, 400, 500
(100 lt 400) spurious retransmission
30
Senders Response without Eifel
cwnd16 ssthresh32
D1

D16
Time-out
D1
cwnd1 ssthresh8
31
Senders Response without Eifel
cwnd16 ssthresh32
D1

D16
Time-out
A2
D1
cwnd1 ssthresh8
32
Senders Response without Eifel
cwnd16 ssthresh32
D1

D16
Time-out
A2
D1
cwnd1 ssthresh8
D2
cwnd2 ssthresh8
D3
33
Senders Response with Eifel
cwnd16 ssthresh32
100
D1, 100, 500

D16, 100, 500
Time-out
ts_first_rexmit 400, cwnd 16, ssthresh32
D1, 400, 500
400
cwnd1 ssthresh8
34
Senders Response with Eifel
cwnd16 ssthresh32
100
D1, 100, 500

D16, 100, 500
Time-out
ts_first_rexmit 400, cwnd 16, ssthresh32
501
A2, 501, 100
D1, 400, 500
400
cwnd1 ssthresh8
401
35
Senders Response with Eifel
cwnd16 ssthresh32
100
D1, 100, 500

D16, 100, 500
Time-out
ts_first_rexmit 400, cwnd 16, ssthresh32
501
A2, 501, 100
D1, 400, 500
400
cwnd1 ssthresh8
(100 lt 400) spurious retransmission detected
401
Restore cwnd 16, ssthresh32
36
Senders Response with Eifel
cwnd16 ssthresh32
100
D1, 100, 500

D16, 100, 500
Time-out
ts_first_rexmit 400, cwnd 16, ssthresh32
501
A2, 501, 100
D1, 400, 500
400
cwnd1 ssthresh8
(100 lt 400) spurious retransmission detected
401
cwnd16 ssthresh32
D17, 401, 500
Restore cwnd 16, ssthresh32
37
Senders Response without Eifel
A1
D1
cwnd16, ssthresh32
D2
A1

D16
A1
A1
3 Dup-ACKs
38
Senders Response without Eifel
A1
D1
cwnd16, ssthresh32
D2
A1

D16
A1
A1
3 Dup-ACKs
cwnd8, ssthresh8 swnd11
D1
39
Senders Response without Eifel
A1
D1
cwnd16, ssthresh32
D2
A1

D16
A1
A1
A5
3 Dup-ACKs
cwnd8, ssthresh8 swnd11
D1
cwnd8, ssthresh8
40
Senders Response with Eifel
A1, 500, 100
cwnd16, ssthresh32
D1, 100, 500
D2, 100, 500

A1, 501, 100
D16, 100, 500
A1, 501, 100
A1, 501, 100
3 Dup-ACKs
41
Senders Response with Eifel
A1, 500, 100
cwnd16, ssthresh32
D1, 100, 500
D2, 100, 500

A1, 501, 100
D16, 100, 500
A1, 501, 100
A1, 501, 100
3 Dup-ACKs
ts_first_rexmit 400, cwnd16,ssthresh32
cwnd11,ssthresh8
D1, 400, 500
42
Senders Response with Eifel
A1, 500, 100
cwnd16, ssthresh32
D1, 100, 500
D2, 100, 500

A1, 501, 100
D16, 100, 500
A1, 501, 100
A1, 501, 100
A5, 501, 100
3 Dup-ACKs
ts_first_rexmit 400, cwnd16,ssthresh32
cwnd11,ssthresh8
D1, 400, 500
100 lt 400, Spurious ReXmit, Restore
43
Senders Response with Eifel
A1, 500, 100
cwnd16, ssthresh32
D1, 100, 500
D2, 100, 500

A1, 501, 100
D16, 100, 500
A1, 501, 100
A1, 501, 100
A5, 501, 100
3 Dup-ACKs
ts_first_rexmit 400, cwnd16,ssthresh32
cwnd11,ssthresh8
D1, 400, 500
100 lt 400, Spurious ReXmit, Restore
cwnd16,ssthresh32
D17, 401, 501

D20, 401, 501
44
Summary

• Spurious time-out
• Without Eifel ssthresh cwnd/2, cwnd 1
• Enter slow start
• With Eifel Restore old cwnd and ssthresh
• Restore previous state
• Spurious fast retransmit
• Without Eifel ssthresh cwnd cwnd/2
• Enter congestion avoidance
• With Eifel Restore old cwnd and ssthresh
• Restore previous state
• Thus Eifel prevents unnecessary reduction in
  load, and spurious retransmissions

45
Some more details

• On detecting a spurious retransmission
• If a single retransmission done
• ssthresh old ssthresh cwnd old cwnd
• If 2 retransmissions done
• ssthresh cwnd/2 cwnd cwnd/2
• If gt 2 retransmissions done
• ssthresh cwnd/2 cwnd 1
• The more spurious retransmissions have occured,
  the more conservative the sender gets
• In either case, sender resumes transmission with
  the next unsent PDU

46
Thus, the Eifel algorithm..

• Enhancement to TCPs error recovery scheme
• Eliminates retransmission ambiguity, there by
  solving problems caused by spurious timeouts and
  spurious retransmits
• Can improve end-to-end performance by several
  tens of percent
• Makes TCP truly wireless compatible

47
Other detection strategies

• DSACK based detection
• Forward-RTO

48
DSACK based detection
D1

D16
Time-out
A2
D1
ACK for D1
A16, A1
DSACK for D1, D1 spuriously retransmitted, Resume
previous state
RFC 3708, Using TCP Duplicate Selective
Acknowledgement (DSACKs) and Stream Control
Transmission Protocol (SCTP) Duplicate
Transmission Sequence Numbers (TSNs) to Detect
Spurious Retransmissions
49
DSACK based detection
D1
cwnd16, ssthresh32

D16
Time-out
D1
cwnd1, ssthresh8
Internet Draft The Eifel Detection Algorithm for
TCP ltdraft-ietf-tsvwg-tcp-eifel-alg-03.txtgt, RFC
2883 DSACK
50
DSACK based detection
D1
cwnd16, ssthresh32

D16
Time-out
A2
D1
cwnd1, ssthresh8
cwnd2, ssthresh8
Internet Draft The Eifel Detection Algorithm for
TCP ltdraft-ietf-tsvwg-tcp-eifel-alg-03.txtgt, RFC
2883 DSACK
51
DSACK based detection
D1
cwnd16, ssthresh32

D16
Time-out
A2
D1
cwnd1, ssthresh8
D2
cwnd2, ssthresh8
D3
Internet Draft The Eifel Detection Algorithm for
TCP ltdraft-ietf-tsvwg-tcp-eifel-alg-03.txtgt, RFC
2883 DSACK
52
DSACK based detection
D1
cwnd16, ssthresh32

D16
Time-out
A2
A3
D1
cwnd1, ssthresh8
A4
D2
cwnd2, ssthresh8
D3
cwnd4, ssthresh8
. . .
Internet Draft The Eifel Detection Algorithm for
TCP ltdraft-ietf-tsvwg-tcp-eifel-alg-03.txtgt, RFC
2883 DSACK
53
DSACK based detection
D1
cwnd16, ssthresh32

D16
Time-out
A2
A3
D1
cwnd1, ssthresh8
A4
D2
cwnd2, ssthresh8
D3
Go-back-N
cwnd4, ssthresh8
. . .
Internet Draft The Eifel Detection Algorithm for
TCP ltdraft-ietf-tsvwg-tcp-eifel-alg-03.txtgt, RFC
2883 DSACK
54
DSACK based detection
D1
cwnd16, ssthresh32

D16
Time-out
A2
A3
D1
cwnd1, ssthresh8
A4
D2
cwnd2, ssthresh8
D3
Go-back-N
cwnd4, ssthresh8
. . .
A16, A1
D1 spuriously retransmitted
Internet Draft The Eifel Detection Algorithm for
TCP ltdraft-ietf-tsvwg-tcp-eifel-alg-03.txtgt, RFC
2883 DSACK
55
DSACK based detection
D1
cwnd16, ssthresh32

D16
Time-out
A2
A3
D1
cwnd1, ssthresh8
A4
D2
cwnd2, ssthresh8
D3
Go-back-N
cwnd4, ssthresh8
. . .
A16, A1
D1 spuriously retransmitted

• Detection usually done after error recovery is
  over
• Cant avoid unnecessary go-back-N retransmissions
• Less robust against ACK losses, as DSACKs are not
  repeated
• each duplicate segment is reported in only one
  ACK packet, hence information about that
  duplicate segment will be lost if that ACK packet
  is lost

Internet Draft The Eifel Detection Algorithm for
TCP ltdraft-ietf-tsvwg-tcp-eifel-alg-03.txtgt, RFC
2883 DSACK
56
Forward-RTO-Recovery (F-RTO)

• An RTO indicates
• Loss in the network
• Excessive delay in PDU delivery while outstanding
  PDUs are still in flight
• If RTO is due to delay
• RTO is spurious
• ACKs for non-retransmitted PDUs should come after
  RTO

• Pasi Sarolahti , Markku Kojo , Kimmo Raatikainen,
  F-RTO an enhanced recovery algorithm for TCP
  retransmission timeouts, ACM SIGCOMM Computer
  Communication Review, v.33 n.2, April 2003
• Internet Draft, F-RTO An Algorithm for Detecting
  Spurious Retransmission Timeouts with TCP and
  SCTP,
• draft-ietf-tcpm-frto-01.txt

57
F-RTO based detection
D1
D15

Time-out
D1
A3
D2 is not retransmitted, ACK A3 indicates that
the RTO was spurious (was probably due to delay)
58
F-RTO based detection explained
D1
cwnd 16
ssthresh 32

D16
Time-out
D1
ssthresh 8
cwnd 1
Look for next 2 ACKs after retransmission
59
F-RTO based detection explained
D1
cwnd 16
ssthresh 32

D16
Time-out
D1
ssthresh 8
cwnd 1
A1
If the first ACK does not advance the window,
60
F-RTO based detection explained
D1
cwnd 16
ssthresh 32

D16
Time-out
D1
ssthresh 8
cwnd 1
A1
If the first ACK does not advance the window,
cwnd 1
Enter slow start
61
F-RTO based detection explained
D1
cwnd 16
ssthresh 32

D16
Time-out
D1
ssthresh 8
cwnd 1
A2
If the first ACK advances the window,
62
F-RTO based detection explained
D1
cwnd 16
ssthresh 32

D16
Time-out
D1
A2
ssthresh 8
cwnd 1
Spurious RTO
Valid RTO
or
A2
If the first ACK advances the window,
63
F-RTO based detection explained
D1
cwnd 16
ssthresh 32

D16
Time-out
D1
A2
ssthresh 8
cwnd 1
or
A2
If the first ACK advances the window,
D17
Send 2 more new PDUs
D18
cwnd 8
64
F-RTO based detection explained
D1
cwnd 16
ssthresh 32

D16
Time-out
D1
A2
ssthresh 8
cwnd 1
or
A2
If the first ACK advances the window,
D17
Send 2 more new PDUs
D18
cwnd 8
A
Duplicate ACK/ ACK for new PDU
65
F-RTO based detection explained
D1
cwnd 16
ssthresh 32

D16
Time-out
D1
A2
ssthresh 8
cwnd 1
or
A2
If the first ACK advances the window,
D17
Send 2 more new PDUs
D18
cwnd 8
A2
Duplicate ACKs Valid RTO
cwnd 1
Enter slow start
66
F-RTO based detection explained
D1
cwnd 16
ssthresh 32

D16
Time-out
D1
A2
ssthresh 8
cwnd 1
A3
or
A2
If the first ACK advances the window,
D17
Send 2 more new PDUs
D18
cwnd 8
or
New segments ACKed Spurious RTO
A17
Enter congestion avoidance
67
Conclusion

• Problem
• Spurious timeouts and spurious fast retransmits
• Senders inability to distinguish an ACK for
  original transmission from the ACK for
  retransmission
• Solution
• Eliminate retransmission ambiguity
• Eifel Algorithm
• DSACK based detection
• F-RTO