2006 Ka-Cheong Leung and Victor O.K. Li 1 TCP in Wireless Networks: Issues, Approaches, and...

2006 Ka-Cheong Leung and Victor O.K. Li 1

TCP in Wireless Networks: TCP in Wireless Networks: Issues, Approaches, and Issues, Approaches, and

ChallengesChallenges**

Dr. Ka-Cheong Leung

*Joint work with Professor Victor O. K. Li


Overview

• Review on TCP• Issues of Running TCP on Wireless Networks• Contributions• Taxonomy of Solutions• Overview of Existing Solutions• Further Discussion• Open Research Issues


Review on TCP• TCP

– byte-stream protocol• cumulative ACK with next expected octet number

– credit-based flow control• advertised window set by the destination

• number of unacknowledged data bytes the source can send to the destination


Review on TCP (Cont’d)• Congestion control operations

– number of packets within the Internet kept below the level at which network performance drops significantly

– cwnd set to 1 (maximum segment size) when a new connection is established

– slow start• cwnd set to 1• cwnd incremented by one for each ACK received

– retransmission timer• timer timeout => a segment loss => network congestion• ssthresh set to the half of the amount of outstanding data• slow start until ssthresh• congestion avoidance phase

– cwnd increased by 1 for each RTT


Review on TCP (Cont’d)• Congestion control operations (Cont’d)

– duplicate ACK• data octet number of an arriving segment is greater

than the expected one• destination finds a gap in sequence number space

(sequence hole)• destination sends a duplicate ACK, an ACK with

the same expected data octet number in the cumulative acknowledgement field

• in-order communication channel– reception of duplicate ACK => a segment loss



– fast retransmit• source receives three duplicate ACKs

• inferred loss segment retransmitted immediately

– fast recovery• fast retransmission suggests the presence of mild network

congestion

• ssthresh set to the half of the amount of outstanding data

• cwnd set to ssthresh + number of duplicate ACKs received

• cwnd reset to ssthresh and congestion avoidance triggered when an ACK for a new segment arrives



– popular TCP variants• TCP Tahoe

– slow start, congestion avoidance, and fast retransmit

– for each inferred segment loss

» ssthresh set to half of the amount of outstanding data

» do slow start

• TCP Reno– TCP Tahoe + fast recovery


Issues of TCP• Taxonomy of wireless networks

– infrastructured networks• planned, permanent network device installations• cellular networks and most WLANs

– static infrastructured networks– set up with fixed topology connected to backbone network– wireless host can connect via a fixed point (based station or access point)

• satellite networks– quasi-static or dynamic topology– space segment: comprises of satellites– ground segment: a number of base stations (gateway stations) through

which all communications via long-haul satellite links take place

• terminal handoff– mobile host (MH) moves away from the coverage of its base station– MH hands over its proxy for communication from one base station to

another one


Issues of TCP (Cont’d)• Taxonomy of wireless networks (Cont’d)

– Ad hoc networks• without a fixed topology• direct communication: the receiver is in the transmission

coverage of the sender• indirect communication

– send messages to a host in its transmission coverage– receiving host relays the messages on its way to the destination

• merits: flexibility, more robust• drawbacks:

– more difficult and complex to perform routing– more difficult to control or coordinate proper operation of an ad

hoc network (for activities like time synchronization, power management, and packet scheduling)


Issues of TCP (Cont’d)• Characteristics of wireless networks

– channel contention• signals are broadcast and interfere with each other

• transmissions may fail for concurrent transmissions within the interference range of either sender

• medium access protocol needed for coordination

• TDMA-based multi-hop wireless networks– limit the number of in-flight segments concurrently

– correlated arrivals of data segments and their ACKs lead to contention for the wireless channel


Issues of TCP (Cont’d)• Characteristics of wireless networks (Cont’d)

– signal fading• signals distorted or weakened

– propagated over an open, unprotected, and ever-changing medium with irregular boundary

– some signal may disperse and travel on different paths due to reflection, diffraction, and scattering caused by obstacles

– mobility• infrastructured networks

– protocol required to ensure seamless transition during a handoff– packets may be lost during a handoff

• ad hoc networks– transmission route recomputed to cater for topological changes– effective and efficient routing protocol needed for frequent

topological changes


Issues of TCP (Cont’d)• Characteristics of wireless networks (Cont’d)

– limited power and energy• power source may not be able to deliver power as

much as the one installed in a fixed device

• hard to receive a continuous supply of power

• effective and efficient operations with power management

– minimize the number of transmissions and receptions for certain communication operations

» minimize the number of retransmissions for an energy efficient TCP


Issues of TCP (Cont’d)• Problems for TCP

– random loss• dropped due to signal fading

• non-congestive segment losses not negligible– violate the working assumption of the traditional

congestion control measures for TCP

– congestion control mechanisms react inappropriately by:

» keeping the sending rate of a TCP connection small

» retransmitting some data segments spuriously


Issues of TCP (Cont’d)• Problems for TCP (Cont’d)

– burst loss• may be initiated by signal fading

– prolonged uncontrollable channel interferences• infrastructured networks

– a chain of packets lost due to a handoff event– frequency: size of coverage region and host mobility

• ad hoc networks– host mobility => topological change or network partition– re-routing process can take some time to complete– some packets may be lost during the process– frequency: transmission range and host mobility

• can lead to serial timer expirations– multiple consecutive timer expirations and retransmissions of the same

data segment within a single blackout period– result in a terribly long period of inactivity of the connection (due to

exponential timer backoff) even after the network conditions has restored to normal


Issues of TCP (Cont’d)• Problems for TCP (Cont’d)

– packet reordering• network behaviour where the receiving order of a flow of

packets differs from its sending order• persistent and substantial packet reordering violates the (near)

in-order channel assumption• result in substantial degradation in application throughput and

network performance• causes:

– link-layer retransmission– infrastructured networks

» packets take different routes due to handoff– ad hoc networks

» re-routing due to topological changes


Contributions• Present an overview of recent developments and explore

some open research issues and challenges– survey end-to-end solutions proposed to date

• require no intermediaries to scoop the state of a connection• may require supporting functions implemented at the routers for the

sake of efficiency and performance enhancements

• Give the readers a new angle to view the existing state of the art– classify the surveyed solutions based on the way they tackle the

problems– focus on enhancements that have been implemented in the TCP

clients

• Provide the readers a short tutorials of the surveyed representative solutions


Taxonomy of Solutions


Taxonomy of Solutions (Cont’d)• Congestion detection approach

– measure the current network conditions– determine whether network congestion has actually occurred– choose a proper traffic control strategy to differentiate congestive

issues from the non-congestive ones based on the measured information

• State suspension approach– detect the current network state– decide when communication activity is suspended and when it can

be resumed to avoid non-congestive losses• Response postponement approach

– delay triggering a traffic control response to alleviate the problems in wireless networks

• Hybrid approach– a collection of methods that can be classified by more than 1

approach described above


Overview of Existing Solutions

• Congestion detection approach– TCP-Peach (Akyildiz, Morabito, and Palzzo, 2001)

• deal with adverse effects found in satellite networks with long propagation delays & high link error rates

• dummy segments– low-priority segments with a copy of recently transmitted

data

– probe for the availability of network resources

– successfully delivered dummy segment indicates that:

» unused network resources exist

» transmission rate can be increased accordingly


• Congestion detection approach (Cont’d)– TCP-Peach (Cont’d)

• sudden start– substitute slow start

– aim to open up the congestion window faster

– transmit one dummy data segment for every until (awnd – 1) dummy segments have been sent

» τ: estimated RTT

– increment cwnd by 1 segment upon the receipt of an ACK for a dummy segment

Overview of Existing Solutions (Cont’d)

awnd



• rapid recovery– replace fast recovery– halve cwnd in response to an inferred segment loss– arrival of an ACK for a data segment

» send 2 dummy segments until a total of 2 cwnd segments have been transmitted

– increment cwnd by 1 segment» arrival of an ACK for a data segment» arrival of an ACK for a dummy segment, after

receiving cwnd ACKs for dummy segments




• merit: maintain ACK-clocking when cwnd is smaller than the number of unacknowledged data segments

• drawbacks:– implicitly assumed that more than half of the dummy segments

are lost in transit for a congestive loss event– all dummy segments can be successfully delivered to the

destination for a non-congestive loss event– wastage of network resources since the delivery of dummy

segments does not result in any gain in connection goodput» TCP-Peach+: NIL segments with unacknowledged data in

place of dummy segments– dummy segments are sent at a rate doubled that before a loss

event is conjectured => congestion at routers» TCP-Peach+: no more than 1 NIL segment sent per ACK

– all routers configured to implement priority-based scheduling



• Congestion detection approach (Cont’d)– TCP-Probing (Lahanas and Tsaoussidis, 2002)

• sender-side solution– aim to enhance performance against random loss and burst loss

• use of probing devices– determine whether network congestion has occurred when a

segment loss is inferred• A-TCP

– invoke a probing cycle upon receiving 3 duplicate ACKs or a retransmission timer expiration

– probe segments sent until the ACKs of a pair of probes are received within the specified time period

– recovery process depends on the status of network congestion– drawbacks: costly to perform and respond slowly to non-

congestive loss• SP-TCP

– avoid triggering more than 1 probing cycle in small time interval




• Congestion detection approach (Cont’d)– TCP Westwood (TCPW) (Casetti et al., 2002)

• sender-side solution• idea: adjust the size of the congestion window upon an inferred

segment loss by monitoring the rate of acknowledging data• upon each ACK arrival

– use the amount of new data acknowledged by that ACK to update the estimate for the available bandwidth of the connection

• when taking congestion control– ssthresh assigned as:

» (estimated available BW) x (minimum RTT) / (segment size)• merit: decouple congestion control from error control• drawbacks:

– some unfriendliness to TCP Reno– overstate the available bandwidth with the presence of ACK compression

» TCP Westwood+: bandwidth sample computed every RTT instead of with each ACK arrival to eliminate the high frequency components contained in the bandwidth samples



• Congestion detection approach (Cont’d)– TCP Veno (Fu and Liew, 2003)

• sender-side refinements on TCP Reno– deal with random loss

• estimate the backlog accumulated along the communication path of the connection

– measured backlog < threshold => no congestion» inferred segment loss as a random loss

• two refinements– congestive loss inferred

» cwnd increased by 1 segment every 2 RTTs instead of each RTT– random loss inferred

» fast retransmit: ssthresh set to 0.8 cwnd instead of 0.5 cwnd• drawbacks:

– performance improvement fades with high random loss rate– fail to deal with multiple segment losses in the same congestion window– may not work well in ad hoc networks

» backlog estimation sensitive to RTT oscillation due to route change



• Congestion detection approach (Cont’d)– TCP-Jersey (Xu, Tian, and Ansari, 2004)

• follow same idea as TCPW to observe the rate of data acknowledged by ACKs– simpler estimator for the available bandwidth

• adopt slow start, congestion avoidance, and fast recovery from TCP Reno• use explicit retransmit instead of fast retransmit

– simply perform a segment retransmission• congestion warning (CW)

– router marks congestion experienced (CE) bit in the IP header of all packets when the average queue length > threshold

– destination echoes the congestion information by setting the explicit echo (ECE) bit of all segments until it receives a segment with the congestion window reduced (CWR) bit set

• ACK arrival– estimate the available bandwidth and compute the optimal size of the congestion

window when not run for 1 RTT– no congestion warning => proceed similarly as TCP Reno– congestion warning

» apply rate control procedure first to set the size of the congestion window based on the computed available bandwidth

» follow the congestion control measures as without congestion warning• drawbacks: aware of CW scheme, fail to handle burst loss



• Congestion detection approach (Cont’d)– JTCP (Wu and Chen, 2004)

• use jitter ratio to determine whether an inferred segment loss congestive or non-congestive

• fast recovery triggered only when:– inferred congestive loss is detected– preceding fast recovery carried out at least 1 RTT ago

• immediate recovery– set ssthresh as D cwnd, where 0.5 < D ≤ 1

• retransmission timer expires– congestive loss: slow start– non-congestive loss: fast retransmit and fast recovery

• drawbacks:– insert and process timestamps– unable to handle burst loss satisfactorily



• Congestion detection approach (Cont’d)– TCP-Casablanca (Biaz and Vaidya, 2005)

• apply a simple biased queue management scheme

• discriminate congestion losses from random losses

• idea: de-randomize congestion losses– distribution of congestive losses differs from random losses

• label 1 “out” segment for every k segments, else “in” segments

• router experiences congestion– drop “out” packets before dropping “in” packets

• dropping sequence will show correlated losses if the lost packets are dropped due to network congestion



• Congestion detection approach (Cont’d)– TCP-Casablanca (Cont’d)

• extended from TCP Newreno– same as TCP Reno except that fast recovery exits only when all sent data

segments are acknowledged before it is entered

• destination uses a simple loss discriminator function to diagnose whether a loss is congestive or not

• non-congestive loss– mark a duplicate ACK with ELN so that the source does not halve the

size of the congestion window

• merit: identify congestive losses with more than 95% accuracy and non-congestive losses with more than 75% accuracy

• drawbacks:– require participating routers to have a differential packet dropping policy– inferior performance with respect to other TCP-friendly flows since

“out” segments are dropped in advance of any other segments



• State suspension approach– Freeze-TCP (Goff et al., 2000)

• improve performance with frequent disconnections• receiver-side solution (for a mobile receiver)• continuously monitor the signal strength of wireless antennas• detect any impending handoffs• about 1 RTT before a handoff

– send some zero window advertisements (ZWAs) to force its peer (sender) into the persist mode

– ZWA piggybacked into an ACK

• receive a ZWA from the receiver– persist mode: freeze all retransmission timers and the size of the congestion

window– send zero window probes (ZWPs) with inter-probe time being backed off

exponentially

• receive a destination’s response with a positive advertised window size– exit from the persist mode– resume its transmission as normal



• State suspension approach (Cont’d)– Freeze-TCP (Cont’d)

• drawbacks:– must be aware of mobility and need some cross-layer

information exchanges– need to predict when a disconnection is expected to

happen– fail to predict and detect an upcoming disconnection event

if it happens at a wireless link along the transmission path– resumed transmission rate may be set inappropriately– can only avoid performance degradation due to

disconnections» fail to avoid and identify occasional segment losses

because of signal fading



• State suspension approach (Cont’d)– ILC-TCP (Chinta, Helal, and Lee, 2003)

• sender-side solution (for a mobile sender)– prevent performance degradation due to temporary

disconnections

• idea: control decision based on the state information– link layer: link state (good or bad)– network layer: IP-level handoff started or completed

• upon a timer expiration– both link and network layers stable => network

congestion => take regular congestion control measures– otherwise: freeze connection state until both layers

become stable



• State suspension approach (Cont’d)– TCP-Feedback (Chandran et al., 2001)

• improve the performance for route failures• route disruption detected

– failure point transmits a route failure notification (RFN) packet to the source

– each immediate mobile host invalidates the route» alternative route exists: reroute packets and discard the RFN packet» otherwise: relay the RFN packet to the source

• source receives the RFN packet– bring the TCP connection to the snooze state until the route failure

timeout or a route reestablishment notification (RRN) packet received• new route learnt by an immediate mobile host

– send an RRN packet to the source• merit: able to handle route disruption at any wireless link• drawbacks:

– burst injection– resumed transmission rate may be set inappropriately



• State suspension approach (Cont’d)– ELFN (Holland and Vaidya, 2002)

• similar to TCP-Feedback• differences with TCP-Feedback

– ELFN relies on the route failure messages for dynamic source routing (DSR) to notify a source about link and route failures

– no route maintenance or invalidation at immediate hosts– no need for any immediate hosts to send or forward a

RRN packet to a source to re-activate a suspended connection

» source probes the network periodically for re-connection



• State suspension approach (Cont’d)– TCP-DOOR (Wang and Zhang, 2002)

• detect route changes through out-of-order events• out-of-order data/ACK detection

– insert the TCP packet sequence number and ACK duplication sequence number, or current timestamps, into each data and ACK segment, respectively

• temporarily disable congestion control– source keeps its state variable unchanged for a time period

• instant recovery during congestion avoidance– source recovers immediately to the state before the

congestion response invoked within a time period• drawbacks:

– transmission rate may be set inappropriately after a route change– fail to perform well in a congested network environment with substantial

persistent packet reordering



• Response postponement approach– DelAck (Altman and Jiménez, 2003)

• use of delayed ACK techniques to improve performance in multi-hop wireless ad hoc networks

• receiver-side solution– reduce channel contentions among data segments and ACKs of the same

TCP connection– as a side-effect to reduce performance degradation due to packet

reordering• idea: delay acknowledging the arrivals of data segments and reduce

the number of ACKs sent to the source– generate an ACK for every d data segments or the first unacknowledged

data segment has been received for a certain time period (e.g. 0.1 s)• merit:

– reduce the connection overhead and hence the channel contentions• drawbacks:

– d is orthogonal to the segment sequence number in general, but DelAck sets it to increase with the sequence number

– burst injection due to delayed acknowledgement



• Response postponement approach (Cont’d)– TCP-ADA (Singh and Kankipati, 2004)

• receiver-side solution, similar to DelAck• postpone acknowledgement for a time period

– defer sending an ACK of a segment for βΔ» Δ is an EWMA of the inter-segment arrival time

– deferment period restarted every time data segment arrives

– ACK sent when deferment timer expires or maximum deferment period is reached

• drawbacks:– send bursts of new segments once every RTT– significant drop in throughput if ACKs are lost



• Response postponement approach (Cont’d)– TCP-DCR (Bhandarkar et al., 2005)

• sender-side solution– meliorate the TCP robustness to non-congestive events

• idea: delay a congestion response for a time interval after the first duplicate ACK is received

• set the time interval as RTT to deal with packet reordering due to link-layer retransmissions

• merit:– perform significantly better than TCP with SACK and TCPW in the

presence of channel errors

• drawback:– chosen period of deferment highly dependent on RTT

» several retransmissions of the same packet can be delayed longer than one RTT




• Hybrid approach– ATCP (Liu and Singh, 2001)

• resolve problems with TCP in ad hoc networks– high bit error rates, frequent route changes, network partitions,

and packet reordering

• idea: introduce a layer (ATCP layer) between TCP and IP at the sender’s protocol stack

– ATCP layer: monitor the current TCP state and spoof TCP from triggering its congestion control mechanisms inappropriately

• apply ECN and ICMP to sense the onset of network congestion and integrity of the transmission path



• Hybrid approach (Cont’d)– ATCP (Cont’d)

• four states– normal: do nothing and transparent– congested: take congestion behaviour of TCP– loss: put TCP in the persist mode and send unacknowledged data

segments– disconnected: place TCP into the persist mode and send probes to the

destination; slow start is invoked when leaving the state

• merit:– able to handle most of the problems relating to wireless networks

• drawbacks:– inefficient in using the available bandwidth with frequent route changes

and network partition– aware of and implemented with ECN– do not allow a source to send new data segments in the loss state



Further Discussion• Congestion detection approach

– either use probes or information stored in the data segments and ACKs to detect the congestion conditions

– TCP-Peach and TCP-Peach+• send low-priority segments to quickly seize the available bandwidth

– TCP-Probing• transmit probes to detect whether the network is congested based on

the estimated RTT

– TCPW, TCP Westwood+, TCP Veno, TCP-Jersey, and JTCP• estimate the network congestion level based on the spatial and

temporal information carried by the ACKs

– TCP-Casablanca• infer the network congestion status based on the ratio of the marked

segments being dropped by the network


Further Discussion (Cont’d)• Congestion detection approach (Cont’d)

– merit:• generally able to distinguish between congestive

loss and non-congestive random loss• determine the appropriate traffic control measure

strategy to improve the TCP performance

– drawbacks:• generally fail to gracefully handle multiple segment

losses in the same congestion window as most approaches extended from TCP Reno

• no specific mechanisms to avoid burst loss due to temporary disconnections


Further Discussion (Cont’d)• State suspension approach

– utilize the state information as well as route failure and restoration notifications

– decide whether the communication activity of a connection is suspended or resumed

– Freeze-TCP• use the signal strength information to infer the occurrence of a temporary

disconnection

– ILC-TCP• freeze the communication whenever either link or network errors are

experienced

– TCP-Feedback and ELFN• stop the communication activity when a route failure notification is received• resume communication after a route is established for a suspended connection

– TCP-DOOR• temporarily disable congestion control or perform instant recovery during

congestion avoidance after detecting an out-of-order event


Further Discussion (Cont’d)• State suspension approach (Cont’d)

– merit:• successful at suspending any congestion control measures and

stopping further segment losses when a temporary disconnection is encountered

• TCP-DOOR: alleviate some performance problems caused by packet reordering

– drawback:• fail to deal with occasional random losses due to transit, short-

lived link errors (from signal fading)

• TCP-DOOR: no mechanisms to deal with non-congestive losses


Further Discussion (Cont’d)• Response postponement approach

– defer taking any traffic control measures to gather more network information to see if the decision needs to be changed

– DelAck and TCP-ADA• delay the issuance of an ACK• reduce the load of the control traffic and thus channel contention• as a side-effect to deal with packet reordering

– TCP-DCR• postpone triggering the congestion control response to a newly

received ACK– merits: with the presence of packet reordering

• able to reduce spurious retransmissions• maintain a larger congestion window

– drawbacks:• fail to clock out traffic during the deferment of congestion response

(except for TCP-DCR)• no mechanisms to deal with non-congestive losses


Further Discussion (Cont’d)• Hybrid approach

– use ECN information and source quench messages to detect the occurrence of network congestion

– utilize the destination unreachable messages to detect temporary disconnections

– merit:• able to handle most of the problems (random loss, burst loss,

and packet reordering) relating to wireless networks– drawbacks:

• inefficient in using the available bandwidth with frequent route changes and network partition

• aware of and implemented with ECN• do not allow a source to send new data segments in the loss

state



Open Research Issues• Integrated solutions for all types of wireless

problems– except for ATCP, none of the surveyed solutions can

deal with all of the aforementioned problems– ATCP

• loss state: stop transmitting new data segments until a new ACK arrives

• presence of persistent packet reordering– block the regular new data segment transmissions– reduce the connection goodput– shrink the battery lifetime of a wireless host due to unnecessary

retransmissions


Open Research Issues (Cont’d)• Improved discriminator for wireless problems

– problem mistaken• improper traffic control may exacerbate the problem

(congestive loss as non-congestive)

– need an efficient and effective discriminator to correctly identify congestive loss, random loss, burst loss, and packet reordering

• Theoretical basis for an efficient and optimal scheme to handle the identified non-congestive losses

Date post:	20-Dec-2015
Category:	Documents
View:	214 times
Download:	0 times

2006 Ka-Cheong Leung and Victor O.K. Li 1 TCP in Wireless Networks: Issues, Approaches, and...

Documents