Transcript Slide 1
University of Berne
Institute of Computer Science and Applied Mathematics – IAM/RVS
TCP Issues in Mobile
IP Networks
Ruy de Oliveira
December 05, 2001
Topics addressed
Brief review on TCP algorithm
Challenges for TCP under mobile environment
Main proposed approaches for cellular net.
Requirements on mobile multi-hop networks
Some proposals for multi-hop environment
Conclusions and outlooks
TCP review
TCP has been designed to work on wired networks
• Negligible medium loss (low BER)
Under loss it starts probing the net at lower rate by
shrinking its congestion window (CWND)
• Slow Start (timeout)
exponential back off (RTO)
• Congestion Avoidance
• Fast retransmit and recover (3 dacks)
Receiver window (RW) limits the maximum rate of
the sender
• Upon receiving a RW set to zero, sender enters into
“persist mode”
TCP under mobile environment
In mobile networks, pck losses refer to:
• Congestion within wired network
• Non-negligible wireless losses (high BER)
• Disconnection (Handover, fading, etc)
As TCP does not discriminate such losses, it can
waste bandwidth by dropping its CWND when
• A pck loss occurs in the wireless link
• A fast handover takes place towards a cell with enough
bandwidth
Serial timeouts
Dealing with TCP in mobile IP
The main techniques used to get over TCP behavior
in mobile networks include:
– To split the e2e connection into two, namely wired and
wireless connection
– To push the sender into “persist mode” during handover by
either the
• Base Station (BS) (or another intermediate node) or
• Mobile Host (MH) (by predicting imminent disconnect.)
– To improve local wireless retransmission
– To speed up the TCP recovery after a handover
TCP approaches for cellular network
I-TCP
Snoop
M-TCP
Delayed duplicate acks (dacks)
EBSN
WTCP
Freeze-TCP
TCP-probing
Fast retransmit
Indirect-TCP (I-TCP)
It splits the e2e connection into two parts:
The wireless connection can even use another
transport protocol that suits wireless medium
During handover pcks from FH are cached at old BS
to be transferred to the new one
It’s backward compatible with fixed network
Indirect TCP operation
I-TCP drawbacks
Maintains no e2e TCP semantics
• BS acknowledges (ACK) pcks to the sender
• It requires cooperation of application layer to provide
reliability
The BS can run out of buffer
High processing at BS
Latency to transfer state information can be
prohibitive
Snoop Protocol
Changes are restricted to BS and optionally to MH
as well
E2e TCP semantics is preserved
A (snoop) layer is added to the routing code at BS
which keep track of pcks in both directions
Pcks meant to MH are buffered at BS and, if
needed, retransmitted in the wireless link
It’s robust in dealing with multiple pck losses in a
single transmission window
Snoop Protocol functioning
Snoop Protocol drawbacks
Recovery from handover can be slow due to
considerable state information to be handed over
Under long disconnection, sender times out
Encrypted traffic cannot be handled
M-TCP Protocol
Also splits the connection into two
Unlike I-TCP, it maintains e2e TCP semantics
Under long disconnection pushes the sender into
“persist mode”
It avoids frequent transferring of state information
during handover
It’s appropriated for environment with high cells
switching
M-TCP Protocol operation
M-TCP Protocol disadvantages
When sender transmit occasionally only, it will time
out as the SH-agent does not send last ACK
Some retransmission overhead
High processing at SH
Considerable complexity
Encryption is not possible
Reliability issues
TCP approaches cellular network
I-TCP
Snoop
M-TCP
Delayed duplicate acks
EBSN
WTCP
Freeze-TCP
TCP-probing
Fast retransmit
Approaches comparison
I-TCP
Events/
feature
E2e
semantics
no
Handle high yes
BER
Snoop
M-TCP
Delayed
dacks
EBSN
WTCP
FreezeTCP
TCPprobing
yes
yes
yes
yes
yes
yes
yes
yes
yes*
yes
yes
yes
no
yes
Long
disconnec.
may run
out buffer
no
yes
no
no
no
yes
yes
Freq.
disconnec
handov.
costly
no
may be
costly
no
no
no
yes
yes
Req. interm
node TCP
mode
yes
yes
yes
no
yes
yes
no
no
Handle
encryption
no
no
no
yes
no
no
yes
yes
power
saving
no
no
yes
no
no
no
yes
yes *
Mobile multi-hop (Ad hoc) networks
Mobile multi-hop = mobile Ad hoc = Manet
This wireless framework is “wired infrastructure”
independent
Each node is both end-user and router
It’s appropriate for environment where wired network
cannot be used or is not desired
TCP challenges in manet networks
All those met in Cellular networks (1-hop)
Environment under high route failures
• Frequent routing changes
• Partitions
Multi-path routing needs to be considered
Power saving awareness is extremely necessary
CWND may not represents actual available BW
(route dependent)
Manet scheme
Approaches for TCP within manets
To
lead the sender into “persist mode” or a similar one
fix the RTO under route failure
make use of feedback information
rely on cooperation from network and link layers
improve link protocol recovery strategy
Some proposals
TCP-F
ELFN-based approach
Fixed RTO
ATCP
TCP-F
Based on feedback scheme
Sender to distinguish route failure from net. cong.
Sender enters snooze state when receives RFN
It resumes transmission when receives a RRN
Lack of RFN or RRN makes it performs like std TCP
ELFN-based approach
Employs the concept of Explicit Link Failure
Notification (ELFN) techniques
Via ELFN sender is told about link and route failures
ELFN carried by routing protocol itself (piggy-back)
Upon receiving an ELFN TCP disables cong. control
• Instead it enters a “stand-by” mode timers frozen
• Starts probing the network
Retransmission resumes at “full rate”
Routing protocol (DSR) staled cache problem
degrades performance significantly
Fixed RTO
The exponential back off algorithm is disable so the
sender retransmits at regular intervals
A 2nd RTO happening, indicates route loss
The scheme was evaluated for two on-demand
(AODV, DSR) and one proactive (ADV) routing
algorithms
• On-demand ones performed well
• Proactive didn’t experience improvement with this
approach
This approach is only feasible for wireless networks
ATCP
Std TCP is not modified Interoperability
Defines ad hoc layer to work between layers 3 and 4
ECN and ICMP “Dest. Unreach.” signaling are used
• ECN congestion
• ICMP router failure (partition or re-computation)
ATCP spoofs TCP to obtain the following behavior:
•
•
•
•
High error Simply retransmit pck from TCP buffer
Route update delay Stop/resume with new CWND
Transient partition idem
Multi-path routing invoke CC
ICMP messages might not reach the sender
ATCP
Based on network feedback atcp puts sender into:
• Persist mode
• Congestion control mode
• Retransmit mode
ATCP State transition at sender
AD Hoc approaches
Event/feature
TCP-F
ELFN-based
Fixed RTO
ATCP
Pck loss due
to high BER
Not handled
Not handled
Not handled
Retransmit without
invoking CC
Route changes
RRN freezes
or network
sender state
partition
ELFN freezes
sender state
Upon 2nd
timeout RTO
is fixed
ICMP message
puts sender in PM
Pack
reordering
Not handled
Not handled
Not handled
Done by ATCP
layer
Congestion
Not handled
Not handled
Not handled
Via ECN TCP CC
invoked quickly
CWND
Old CWND
used
Old CWND
used
Old CWND
used
Reset for each new
route
wired net.
Req. routing
Requires routing
Requires ECN use
Not smoothly
interoperation algorithm aware algorithm aware
on wired network
Power saving
Not handled
Not handled
Not handled
Not handled
Conclusions and outlooks
Any TCP improvements need consider interoperab.
Power saving awareness is essential
Cooperation among protocol layers seems to be
unavoidable
Further investigation on CWND on resuming
BS tends to be part of encryption scheme
Ad hoc networks (multi-hop)
• TCP performance is Highly dependent on routing pr.
» Geographical-based location protocol seems to be useful
• Link layer strategies to play a key role (high BER)
• Longer periods of disconnection is highly likely