Chapter 6
SACK-TCP, TCP Extensions,
Wireless TCP
Professor Rick Han
University of Colorado at Boulder
[email protected]
Announcements
• Read Sections 6.1 - 6.4, Skip 6.5
• Programming Assignment #1 emailed back by
tonight
• HW #4 on Web by Friday evening
• Programming Assignment #2 due April 2
• Midterm
•
Still grading, hand back April 2
• Next, finishing up TCP
Prof. Rick Han, University of
Colorado at Boulder
Recap of Previous Lecture
• TCP Adaptive Retransmission
•
•
•
Initially,
•
•
Compute exponentially weighted moving average (EWMA) of
RTT
RTO = 2* (EWMA of RTT)
•
•
Compute EWMA of RTT and absolute deviation
RTO = (EWMA of RTT) + 4*(EWMA of abs dev)
•
•
If no retransmission, recompute RTO and EWMA’s
If retransmission, back off RTO = RTO * 2, don’t recompute
EWMA’s
Revision
Retransmission Ambiguity Revision
Prof. Rick Han, University of
Colorado at Boulder
Recap of Previous Lecture (2)
• TCP Congestion Control
•
•
•
•
W = min (FW, CW)
“Slow” Start grows CW exponentially from CW=1
If timeout,
• Reset CW to 1, ssthresh = CW/2
• Slow Start again, up to ssthresh
• Additive Increase after ssthresh
• TCP Tahoe
If 3 duplicate ACK’s
• Retransmit immediately before timeout =
Fast Retransmit
• Drop CW = CW/2 (don’t slow start to CW=1) =
Fast Recovery
Prof. Rick Han, University of
• TCP Reno
Colorado at Boulder
Timer Granularity
• Many TCP implementations set RTO in
multiples of 200,500,1000ms
– Avoid spurious timeouts – RTTs can vary
quickly due to cross traffic
– Make timers interrupts synchronized
and therefore efficient
• At startup:
– Pick a very conservative value for RTO
on order of seconds
Prof. Rick Han, University of
Colorado at Boulder
Delayed ACKS
• Problem:
– In interactive programs like telnet, you could
send a TCP packet for each character you type
– Overhead becomes high:
• You send 1st character, get an ACK from remote host
• Remote host sends echo of character for you to draw
on screen, gets an ACK from you for echoed data
• 4 segments exchanged
• Solution:
– Wait 200ms before ACK’ing
• Can piggyback echo of data with ACK of 1st character
=> only 3 segments exchanged
– Must ACK every other segment
– Must not delay duplicate ACKs
Prof. Rick Han, University of
Colorado at Boulder
TCP ACK Generation
[RFC 1122, RFC 2581]
Event
TCP Receiver action
In-order segment arrival,
No gaps,
Everything else already ACKed
Delayed ACK. Wait max 500ms
for next segment (usually 200ms timer).
If no next segment, send ACK
In-order segment arrival,
No gaps,
One delayed ACK pending
Immediately send single
cumulative ACK
Out-of-order segment arrival
Higher-than-expect seq. #
Gap detected
Send duplicate ACK, indicating seq. #
of next expected byte
Arrival of segment that
Partially or completely fills gap
Immediate ACK if segment starts
at lower end of gap
Prof. Rick Han, University of
Colorado at Boulder
Courtesy: Srini Seshan
SACK-TCP, RFC 2018 & 1072
• Problem: Cumulative ACK’s can’t identify
whether there are multiple losses in a window
– If retransmit only lowest segment, then too slow
(retransmit serially) if there are multiple losses
– If retransmit entire window, then could retransmit
unnecessarily if only one loss
• Solution: use Selective ACK’s
– SACK-TCP (proposed 1988, modified 1996) ACKs
only identify up to 3 non-contiguous blocks of
received data
– Not quite SRP
• Higher throughput than TCP Reno when
multiple segments are lost in a window of data
Prof. Rick Han, University of
Colorado at Boulder
SACK-TCP (2)
• During connection setup, a SACK-TCP sender
sets “SACK-permitted” bit in SYN packet
– Informs receiver that sender is SACK-enabled
• If receiver supports SACK-TCP
– receiver responds to data packets from sender by
extending TCP header with a TCP option called a
“SACK option”
– SACK option identifies non-contiguous blocks of
correctly received data (not the gaps)
• up to 3 such blocks – limited by size of TCP options header,
assumes that TCP timestamp extension is also used
– 3 is enough for “substantial benefit” over TCP Reno
• If more TCP extensions are used, then only enough space to
identify 1 or 2 SACK blocks
Prof. Rick Han, University of
Colorado at Boulder
SACK-TCP (3)
• At sender,
– Each selectively ACK’ed segment has a bit
set
– Only segments with bits not set are
selectively retransmitted
– Since receiver may choose to discard out-oforder segments, then when sender times out,
sender resets all SACK bits to zero
– Try to preserve prior congestion control
strategies
Prof. Rick Han, University of
Colorado at Boulder
Interaction of TCP Flavors
• In addition to TCP Reno, Tahoe, and SACK
– also have NewReno and Vegas
• SACK-TCP implemented in Linux 2.1.90 and
later, Microsoft Windows 98/2000, Sun
Solaris 2.6 and later
• How do Reno and SACK interact?
– If SACK sender and Reno receiver, then Reno
receiver ignores “SACK-permitted” and just
sends cumulative ACKs
– If Reno sender and SACK receiver, then
receiver just sends cumulative ACKs without
SACK options
Prof. Rick Han, University of
Colorado at Boulder
TCP Extensions/Options
• Examples: Timestamp, Large windows, SACK-TCP
• Implemented by extending TCP header > 20
bytes:
– Hdr len/data offset field in TCP header specifies hdr
size in 4-byte words (>5 if option present)
– After 1st 20 bytes, the 1st byte of options field
specifies the “kind” of option
•
•
•
•
•
2 = “Maximum Segment Size”
3 = “Window Scale”
4 = “SACK-permitted”
5 = “SACK-option”
8 = “Time stamp extension”
– Following bytes contain length (optional) and option
data
Prof. Rick Han, University of
Colorado at Boulder
TCP Timestamp Extension,
RFC 1323
• Used to improve timeout mechanism by
more accurate measurement of RTT
• When sending a packet, insert current
timestamp as an option
– 4 bytes for seconds, 4 bytes for microseconds
• Receiver echoes timestamp in ACK
– Actually will echo whatever is in timestamp
• Advantages: sequence #’s are qualified by
time and hence unique:
– Removes retransmission ambiguity
– Protects against wraparound
– Accurate RTTProf. Rick Han, University of
Colorado at Boulder
TCP Large Windows Option,
RFC 1072, 1323
• Delay-bandwidth product for Long
Fat Network (LFN – elephant!) => 16bit window (65536 bytes) is too small
• Scaling factor on advertised window
– Specifies how much to scale (multiply by
left shift) a window size
– Scaling factor exchanged during SYN
connection setup
Prof. Rick Han, University of
Colorado at Boulder
TCP Problems Over Wireless
Links
• Wireless links are inherently error-prone
– Fades, interference, attenuation
– Errors often happen in bursts
• TCP reacts to corruption-based packet loss
by slowing down retransmission and
constricting congestion window
– This is wrong reaction, because there’s no
congestion on wireless link, only bit corruption
• Performance degradation
– Should continue to retransmit and send new
data, rather than slow down and constrict CW
Prof. Rick Han, University of
Colorado at Boulder
TCP Problems Over Wireless
Links (2)
• Assume following topology: A <-> G <-> B
– Gateway G separates the wired and
wireless worlds
– Node A in wired Internet wants to send to
node B in wireless world through gateway G
– Aggregate the effects of wired Internet
into one connection between A and G
– Aggregate the effects of wireless world
into one link between G and B
Prof. Rick Han, University of
Colorado at Boulder
Congestion vs. Wireless Bit
Corruption
Router
Computer 1
Computer 2
Loss  Congestion
3
2
22
1
0
Loss  Congestion
Wireless
Prof. Rick Han, University of
Colorado at Boulder
Proposed Solutions
• End-to-end protocols
– Selective ACKs, Explicit loss notification
• Split-connection protocols
– Separate connections for wired path and
wireless hop
• Reliable link-layer protocols
– Error-correcting codes
– Local retransmission
Prof. Rick Han, University of
Colorado at Boulder
End-to-End Solutions
• Improve implementations ofTCP at endpoints
– Help sender distinguish between congestion and
corruption via flags
• Network routers could set Explicit Congestion
Notification Bit (ECN) in IP header
• And/or Wireless endpoint could signal Explicit Loss
Notification (ELN) to indicate wireless loss/low SNR
– ACKs include flag indicating either congestion or
corruption loss
Wired link
Wireless link
Prof. Rick Han, University of
Colorado at Boulder
End-to-End: Selective Acks
4 3
Wireless loss
X2
6 5
Correspondent
Host
Cumative
ACK 1
Base Station
Can set ECN
CACK 1,
SACK 3
CACK1,
SACK 3-4
CACK1,
SACK 3-5
1
Mobile Host
CACK 1,
SACK 3-6
Since ECN is not set in ACKs, then Sender infers
corruption caused loss of packet 2, not congestion
Drawbacks: have to modify IP to allow ECN, have to modify TCP
to put ECN feedback in ACK
Split Connection Solutions
• Split connections
– Run TCP over wired link, Run TCP or another
protocol over wireless link
– Let TCP tailored for wireless handle wireless
corruption, let wired TCP handle wired
congestion
Wired link
Wireless link
Prof. Rick Han, University of
Colorado at Boulder
Split Connection
A-D belong to packet 1
3 2
X
Correspondent
Host
ack 0
X
1
D C B
Mobile Host
Base Station
ack 0
cack A
A
cack B
cack B,
sack D
Split Connection Drawbacks
• Consistent wireless losses will back up and
overflow the intermediate node, backing up to
sender, causing congestion backoff though there’s
no congestion
• Violates end-to-end semantics
– When sender receives an ACK, sender believes ACK’ed
data was correctly delivered to destination
– No longer true: intermediate node could fail
Wired link
Wireless link
Prof. Rick Han, University of
Colorado at Boulder
Link Layer Solutions
• More aggressive local rexmit than TCP
– Bandwidth not wasted on wired links
• Drawbacks: Adverse interactions with transport layer
– Timer interactions
– Interactions with fast retransmissions
– Large end-to-end round-trip time variation
• FEC does not work well with burst losses
Wired link
Wireless link
ARQ/FEC
Prof. Rick Han, University of
Colorado at Boulder
Hybrid Solution: TCP Snoop
• Transport-aware link protocol
• Modify base station:
– To cache unACKnowledged TCP packets
– To suppress duplicate ACKs back to sender while
performing local retransmissions
• Key advantages:
– No transport termination or TCP code in base
station
• Preserves end-to-end semantics
• Only soft state in basestation: easy to migrate, loss of
soft state merely returns TCP to its bad wired-overwireless performance
– Fixed and mobile endpoints don’t require any
modifications Prof. Rick Han, University of
Colorado at Boulder
Snoop Protocol: CH to MH
4 3 2 1
6 5
Correspondent
Host
Snoop Agent
1
Base Station
Mobile Host
Example Courtesy of Srini Seshan
• Snoop agent: active interposition agent
– Snoops on TCP segments and ACKs
– Detects losses by duplicate ACKs and timers
– Suppresses duplicate ACKs from FH sender
Snoop Protocol: CH to MH
Snoop Agent
65 3 2
4 1
Correspondent
Host
Base Station
• Transfer of file from CH to MH
• Current window = 6 packets
Mobile Host
Snoop Protocol: CH to MH
65
Snoop Agent
4
3
2
Correspondent
Host
• Transfer begins
1
Base Station
Mobile Host
Snoop Protocol: CH to MH
4 3 2 1
6 5
Correspondent
Host
Snoop Agent
1
Base Station
Mobile Host
• Snoop agent caches (black) segments that pass
by
• Difference #1 from pure link-layer – does not
add a new header uses existing TCP header to
identify losses
Snoop Protocol: CH to MH
4 3 2 1
6 5
Correspondent
Host
• Packet 1 is Lost
Snoop Agent
3 2 1
Base Station
Mobile Host
1
Lost Packets
Snoop Protocol: CH to MH
5 4 3 2 1
6
Snoop Agent
4 3
2
cack 0
Correspondent
Host
Base Station
Mobile Host
1
Lost Packets
• Packet 1 is Lost
– Duplicate ACKs generated
– Note: technically for TCP, “cack0” should be “cack 1”, since TCP
lists in its cumulative ACK the *next* byte it expects. I’ve kept
“cack 0”, i.e. all packets up to 0 have been received correctly.
Snoop Protocol: CH to MH
6
5 4 3 2 1
Snoop Agent
6
5
1
4
3
2
cack 0
Correspondent
Host
Base Station
cack 0
Mobile Host
1
Lost Packets
• Packet 1 is Lost
– Duplicate ACKs generated
• Packet 1 retransmitted from cache at higher
priority
Snoop Protocol: CH to MH
6
5 4 3 2 1
Snoop Agent
6
5
1
4
3
2
cack 4
Correspondent
Host
Base Station
cack 0
Mobile Host
X
• Duplicate ACKs suppressed
• Difference #2 from pure link-layer – tries to
prevent sender from noticing loss
– Sender may still timeout though – fortunately
timeouts are typically long (500ms+)
Snoop Protocol: CH to MH
6
5
Snoop Agent
6
5 1
4
3
2
cack 5
Correspondent
Host
Base Station
cack 4
• Clean cache on new ACK
Mobile Host
Snoop Protocol: CH to MH
6
cack 4
Correspondent
Host
Snoop Agent
6 51
4
3
2
cack 6
Base Station
cack 5
• Clean cache on new ACK
Mobile Host
Snoop Protocol: CH to MH
Snoop Agent
9
8
Correspondent
Host cack 5
7
Base Station
cack 6
6 51
4
3
2
Mobile Host
• Active soft state agent at base station
• Transport-aware reliable link protocol
• Preserves end-to-end semantics
TCP Snoop Drawbacks
• Base station vendors must modify their
firmware to add TCP snoop agent
– Additional cost and complexity, may eventually be
worth it if wireless data usage becomes
commonplace and TCP performance is still an issue
• Return path for ACKs may not traverse same
basestation as forward data path
– Snoop agent that has soft state for unACKed data
doesn’t see ACKs
– Another snoop agent that doesn’t have the soft
state sees ACK’s but can’t suppress duplicates
Prof. Rick Han, University of
Colorado at Boulder