A Comparison of Mechanisms for Improving TCP Performance over Wireless Links

advertisement
A Comparison of Mechanisms for Improving
TCP Performance over Wireless Links
By: Hari B., Venkata P. et. al.
Presented by: Nitin Bahadur
Nitin Bahadur
Advanced Computer Networks
Spring 2000
How I plan to keep you Awake
• Review of TCP Congestion Control and Wireless issues
• Discussion of techniques presented in the paper
• Evaluation of some techniques
• What is a Handoff ???
• New approaches proposed in recent years
Nitin Bahadur
Advanced Computer Networks
Spring 2000
TCP Congestion Control
• Fast Retransmit
– if three duplicate acks before timeout, retransmit
• Fast Recovery
– no slow start after retransmit
– go directly to half the last successful congestion win.
( Cwin = Cwin/2 )
• Coarse grained Timeouts
Assumes packet losses are due to congestion
Assumes an underlying wired network
Nitin Bahadur
Advanced Computer Networks
Spring 2000
Implications to Wireless Networks
• Wireless losses are different from congestion losses
– weak signal, corruption, incomplete packet, lost bits
• TCP treats both losses similarly
– reduces congestion window size
– degrades performance for wireless
• Coarse grained timeouts are bad for lossy wireless networks
– slower retransmissions
– consistent small window size
– reduced bandwidth !!!!
Nitin Bahadur
Advanced Computer Networks
Spring 2000
Solutions
• Approaches presented in the paper
– Split Connection
– End 2 End
– Link Layer TCP aware
• Other recent ones
Nitin Bahadur
Advanced Computer Networks
Spring 2000
Receiver
sends
ack+for
up no.
to 3ofsets
of non-contiguous
data received
Cumulative
Ack.
Seq.
packet
causing the ack.
0 3
0123456789
0
2
3
0-9
1, 3
0-2
1
4-6
8-9
Base Station
Sender can determine which packets were lost before
timeout occurs or 3 duplicate acks are received
Smart assumes no packet reordering on wired link
Nitin Bahadur
Advanced Computer Networks
Spring 2000
Split Connection Schemes
• Divide TCP connection into 2 connections…..isolate wired
network from wireless network
• Use SACK or SMART for performance enhancement
• Loss of end-end TCP model
• Limited buffering available at base station
• Timeouts on wired TCP due to retransmissions on wireless TCP
• Problems in handoff as it now involves 2 TCP connections
TCP II
Wireless Link
Nitin Bahadur
TCP I
Wired N/w
Advanced Computer Networks
Spring 2000
End -End Schemes w/ SMART or SACK
•
•
•
•
Using SMART/SACK, sender can detect multiple losses faster
Faster and efficient retransmit scheme
No need for 3 duplicate acks or coarse timeout
End -End model is maintained
• Still considers wireless losses as due to congestion
• Does invoke congestion control….small congestion window
Nitin Bahadur
Advanced Computer Networks
Spring 2000
E2E w/ Explicit Loss Notification
• Pkt. Loss on wireless link -> Ack. w/ ELN bit set
• Sender retransmits on receiving first (not third) duplicate ack w/
ELN bit set
• Power and time saving !!!!!
• Sender does not invoke congestion control in such cases
large congestion window……even at high rate wireless
losses
How to distinguish b/w congestion and wireless losses ?
Scheme does not detect multiple losses….add SACK/SMART
Nitin Bahadur
Advanced Computer Networks
Spring 2000
LL-SMART-TCP-Aware Scheme
•
•
•
•
Maintain cache of un-acked packets at Home Agent
Use a LL retransmission scheme with finer granularity timeout
Use SMART for efficient retransmissions
Suppress duplicate ack from reaching sender
0 1 2 3
0 3
1,3
Base Station
Layer Violation !!!!!
Bursty losses/slow wireless links lead to TCP sender timeouts
while agent is trying to retransmit
Nitin Bahadur
Advanced Computer Networks
Spring 2000
Effectiveness of LL and E2E schemes
Nitin Bahadur
Advanced Computer Networks
Spring 2000
Handoff Issues
• Mobile hosts (MH) and cell
• Handoff takes place when MH changes Base Station
• Losses due to handoff….During establishing of new route/new
cell registration
• Rerouting through BS or direct routing to MH ???
X
Sender
Nitin Bahadur
Advanced Computer Networks
Spring 2000
Conclusion
The paper presented a taxonomy and comparison of various
approaches
But all approaches have drawbacks…….so none have become
a standard today.
The results presented do not consider losses arising from
congestion…..so are not practical.
How I wish the figures were animated for better understanding !!
Nitin Bahadur
Advanced Computer Networks
Spring 2000
Explicit Bad State Notification (EBSN)
• Base Station sends EBSN message to sender if packets
cannot be transmitted successfully
• Sender changes Timeout based on current RTT
• Timeout is reset to original on receipt of new ack.
• Eliminates unnecessary timeouts
0 1 2
0
EBSN
Nitin Bahadur
Advanced Computer Networks
Spring 2000
Multiple Acks Proposal
• Base Station sends a Partial Ack to sender
• Base station reliably sends packets to mobile client
• Sender does not retransmit/invoke congestion control on
timeout, just discards the Partial Ack
• Receiver sends Complete Ack to sender
• Similar to ELN……but results in excess traffic towards
sender
Nitin Bahadur
Advanced Computer Networks
Spring 2000
Delayed Duplicate Acks (Dupacks)
•
•
•
•
•
TCP - unaware technique, good for encrypted data
Base Station uses a LL retransmission scheme
This scheme uses LL acks…not TCP duplicate acks !!
TCP receiver delays 3rd & other Dupacks
High Priority to LL acks & retransmitted pkts
Receiver cannot distinguish between congestion and
transmission losses…..performance degradation
Nitin Bahadur
Advanced Computer Networks
Spring 2000
Other Proposed Schemes
• Explicit Loss Notification to Receiver (ELNR)
• Explicit Delayed Dupack Activation Notification (EDDAN)
• Wireless Explicit Congestion Notification (WECN)
• Forward Explicit Congestion Notification (FECN)
• Extended Link Failure Notification (ELFN)
• Appropriate Byte Counting
• Loss Predictors
Nitin Bahadur
Advanced Computer Networks
Spring 2000
Download