Decongestion Control

1
Decongestion Control

• By Nabil Samir Sorial
• 27 December 2007

2
Outline

• Congestion Control
• Principles of Congestion Control.
• TCP Congestion Control.
• Fairness.
• Decongestion control
• Introduction
• Online codes
• CHOKe - A stateless active queue management
  scheme for approximating fair bandwidth
  allocation
• Selfish behavior and stability of the internet A
  Game-Theoretic Analysis of TCP
• Benefits
• Design
• Challenges
• References

3
Principles of Congestion Control

• The cause and the costs of congestion
• Three complex scenarios of congestion
• Congestion cost (not fully utilized and poor
  performance)
• Scenario 1 Tow senders, a router with infinite
  buffers
• ?in lt R/2 ? throughput sending rate (?in),
  finite delay
• ?in gt R/2 ? throughput R/2, infinite delay

4
Principles of Congestion Control

• Scenario 2 Tow senders, a router with finite
  buffers
• Packets will be dropped in case full buffer
• If the packet is dropped the sender will
  retransmit it
• Timeout is large enough ? R/3 original data, R/6
  retransmitted data
• Retransmit a packet that not yet lost ? packet is
  forwarded twice, throughput R/4

5
Principles of Congestion Control

• Scenario 3 Four senders, a router with finite
  buffers, multihop paths
• Timeout/retransmission mechanism is used
• If ?in is large, arrival rate of B-D traffic at
  R2 can be much larger than that of A-C traffic
• Buffer at R2 is filled by B-D packets, throughput
  of A-C connection at R2 goes to zero
• The work done by first router is wasted

6
TCP Congestion Control

• TCP uses end-to-end congestion control, IP layer
  provides no explicit feedback to the end systems
  regarding network congestion
• Each sender limits the sending rate as a function
  of perceived network congestion
• AIMD (Additive-Increase, Multiplicative-Decrease)
  
• additive increase increase CongWin by 1 MSS
  every RTT
• multiplicative decrease cut CongWin in half
  after loss
• Saw-toothed pattern

7
TCP Congestion Control

• SS (Slow Start)
• Instead of increasing its rate linearly, TCP
  sender increases its rate exponentially during
  the initial phase (CongWin initialized to 1 MSS)
• Until there is a loss event, the CongWin is cut
  in half and then grows linearly
• Reaction to Timeout event
• Triple duplicate ACKs ? CongWin is cut in half,
  increases linearly
• Timeout ? slow start phase, until CongWin reaches
  one half of it had before timeout, CongWin grows
  linearly (TCP Reno)

8
Fairness

• K TCP connections, passing through a bottleneck
  link with transmission rate R bps
• Fair congestion control if average transmission
  rate of each connection is approximately R/K
• Is TCPs AIMD algorithm fair?
• Different TCP connections may start at different
  times, may have different window sizes at a point
  in time.
• Assume two connections have the same MSS and RTT
• If they have the same window size,then they have
  same throughput

9
Fairness

• The goal is to have the throughputs fall in
  somewhere near the intersection of the equal
  bandwidth line and the full bandwidth utilization
  line
• Point A ? link bandwidth lt R, increase by 1 MSS
  per RTT for both connections (equal increase line
  AB)
• Point B ? link bandwidth gt R, packet loss occurs,
  connection 1 and 2 decrease their windows by
  factor of 2 (point C)

10
Fairness

• We assumed that only TCP connections traverse the
  bottleneck link and that connections have the
  same RTT value
• These conditions are typically not met
• The connections with a smaller RTT are able to
  grab available bandwidth more quickly and will
  enjoy higher throughput than the connections with
  larger RTTs

11
Decongestion control

• Introduction
• Congestion control is fundamental to network
  design to enjoy traffic stability, performance,
  and fairness
• TCP and TCP-like congestion control limit
  transmission rate to explicitly avoid congestion
• Fairness has been achieved by either using fair
  queuing at routers or using common congestion
  control protocol at end hosts
• Fair queuing is expensive to implement and
  congestion control protocol is far from optimal
• Decongestion control relies upon greedy,
  high-speed transmission that will achieve better
  performance and fairness than TCP

12
Decongestion control

• Using efficient, high-speed erasure coding, no
  need to avoid packet loss
• Decongestion control depends on fair dropping and
  erasure-coded data streams that is both efficient
  and fair
• Each sender simply sends as fast as possible. If
  congested routers drop packets in a fair manner,
  each flow will receive its max-min fair
  throughput
• If flows use efficient erasure coding, they will
  achieve goodput almost equal to their throughput,
  fully utilizing network bandwidth.

13
Decongestion control

• Related issues
• Online codesIt is one of efficient, high-speed
  erasure coding that may be used by decongestion
  control protocol
• CHOKe - A stateless active queue management
  scheme for approximating fair bandwidth
  allocationDecongestion control may use CHOKe as
  one of protocols that achieve fairness
• Selfish behavior and stability of the internet A
  Game-Theoretic Analysis of TCPDecongestion
  control uses game-theoretic to show that greedy,
  high-speed transmission will achieve better
  performance

14
Online codes

• Exchanging information over a channel it may lose
  packets, example using UDP over the Internet
• There is a solution FEC (forward-error
  correction) if there is a bound to how many
  packets a channel can lose
• FEC the sender prepares an encoding of the
  original message and sends it over the channel
• The receiver is assured to be able recover the
  original message from the packets that it ends up
  receiving
• The size of the encoded message is bigger than
  the size of the original message to make up for
  the expected losses
• The receiver recovers the original message as
  soon as they receive at least as big a part of
  the encoding as the size of the original message

15
Online codes

• Channel that loses no more than a fraction d of
  the packets, also referred to as a channel of
  capacity 1 d
• There are codes of rate R 1 d for
  transmission over such channels. This means that
  a message of size n blocks can be encoded into
  n/R transmission blocks
• So that any n of the transmission blocks can be
  used to decode the original message
• Elias' codes takes at least O (n log n) time to
  encode and decode
• Tornado codes require linear encoding and
  decoding time
• Online codes are linear encoding/decoding time
  codes, similar to Tornado codes

16
Online codes

• Online code is called the free erasure channel
  that has no constraints on its loss rate
• No fixed-rate erasure code could be used for this
  channel
• Online codes properties local encodability and
  rateless-ness
• Local encodability is the property that each
  block of the encoding of a message can be
  computed independently from the others in
  constant time.
• Rateless-ness is the property that each message
  has an encoding of practically infinite size

17
CHOKe - A stateless active queue management
scheme for approximating fair bandwidth allocation

• The problem is providing a fair bandwidth
  allocation to each of n flows that share the
  outgoing link of a congested router
• The buffer at the outgoing link is a simple FIFO,
  shared by packets belonging to the n flows. CHOKe
  is a simple packet dropping scheme
• CHOKe is queue management algorithm that is
  stateless and easy to implement
• There are two types of router algorithms
  scheduling algorithms and queue management
  algorithms
• Scheduling algorithms can provide a fair
  bandwidth allocation, but it is complex to
  implement

18
CHOKe - A stateless active queue management
scheme for approximating fair bandwidth allocation

• RED (Random Early Detection)
• It is one of queue management algorithms
• Single FIFO to be shared by all the flows, and
  drops an arriving packet at random during periods
  of congestion.
• The drop probability increases with the level of
  congestion
• RED is unable to penalize unresponsive flows
• This is because the percentage of packets dropped
  from each flow over a period of time is almost
  the same.
• Misbehaving traffic can take up a large
  percentage of the link bandwidth

19
CHOKe - A stateless active queue management
scheme for approximating fair bandwidth allocation

• CHOKe
• CHOKe drops more of misbehaving flows packets.
• CHOKe (CHOose and Keep for responsive flows,
  CHOose and Kill for unresponsive flows)
• When a packet arrives at a congested router,
  CHOKe draws a packet at random from the FIFO
  buffer and compares it with the arriving packet
• If they both belong to the same flow, then they
  are both dropped, else the randomly chosen packet
  is left intact
• And the arriving packet is admitted into the
  buffer with a probability that depends on the
  level of congestion
• This probability is computed exactly as in RED

20
CHOKe - A stateless active queue management
scheme for approximating fair bandwidth allocation

• The CHOKe algorithm

21
Selfish behavior and stability of the internet A
Game-Theoretic Analysis of TCP

• This paper attempts to answering the following
  fundamental questionIf network end-points
  behaved in selfish manner, Would the stability of
  the Internet be endangered?
• TCP Game
• Each flow attempts to maximize the throughout it
  achieves by modifying its congestion control
  behavior, using combination of analysis and
  simulation to determine whether the network
  operates efficiently
• TCP Game is defined in which TCP flows in a
  network can adjust their AIMD congestion behavior
• These flows are allowed to freely change and set
  their congestion control parameters, (a, ß),
  where a is the additive increase component and ß
  is the multiplicative decrease component

22
Selfish behavior and stability of the internet A
Game-Theoretic Analysis of TCP

• The Goal is to determine the congestion
  parameters (aE, ßE) at Nash equilibrium, also the
  behavior and efficiency by measuring average
  goodput and the average loss rate
• The Nash equilibrium reflects a balance between
  the gains and the cost related to aggressive
  behavior
• Results from analysis and simulation of TCP Game
• Greed end-point behavior can result in efficient
  network operation with setting of TCP-Reno loss
  recovery in a network of drop-tail routers
• With setting of TCP-SACK loss recovery and RED,
  the result is either low network goodput or high
  drop rates or both, but in using very simple
  dropping algorithms such as CHOKe can restor the
  efficiency

23
Decongestion Control

• Benefits
• Fairness and efficiency
• Senders always transmit at the maximum available
  rate fairness is ensured by dropping policies at
  congested routers
• Tuning the coding rate between sender and
  receiver by dynamically adjust the coding rate
  based on recent throughput rates
• Simplified core infrastructure
• Decongestion control enables significantly
  simpler router designs. Idealized decongestion
  control only requires a fair dropping mechanism
• Erasure coding can reduce the need for queuing in
  the network coding schemes can provide similar
  goodput with coding buffer sizes on the same
  order as router buffers

24
Decongestion Control

• Incentive compatibility
• Decongestion control is more robust to malicious
  behavior due to its time independence
• Senders adjust coding rates based upon reported
  throughput, not individual packet events
• So they are not as sensitive to short-term packet
  behaviors as TCP

25
Design

• Design of decongestion control protocol, Achoo
• Achoo sends erasure-coded packets fast as
  possible between a sender and a receiver
• Packets are labeled with unique, increasing
  sequence numbers
• The receiver periodically acknowledges packet
  reception with information about the rate of
  reception
• All routers implement a fair dropping policy to
  ensure that each flow receives its fair share of
  link bandwidth
• Achoos transmission behavior is controlled by
  two components at the sender the decongestion
  controller and the transmission controller.

26
Design

• Decongestion controller
• All data to be sent is divided into caravans
• Each caravan consists of n fixed-size (1Kb, say)
  data blocks we pick n dynamically
• The role of the decongestion controller is to
  select the appropriate rate of transmission, rate
  of coding, and caravan size
• Caravan size
• The controller starts with a fixed-size caravan
  and begins the transmission loop
• When a caravan is successfully delivered, the
  controller doubles the size of the next caravan
• If after some fixed timeout (likely a function of
  the RTT) there is insufficient data in the socket
  buffer to fill a caravan, the caravan size is
  halved

27
Design

• Type and rate of coding
• Many strategies can be used for each caravan
• Small caravans consist of duplicate data
  (ordinary redundancy)
• Large caravans use rateless erasure codes
• Senders will adjust the rate of coding in
  response to the successful delivery rates
  reported by the receivers
• Transmission rate
• Each physical interface has a maximum achievable
  rate
• The controller determines which flows can put
  that capacity to most effective use the nth flow
  on an interface is given 1/n of the link capacity

28
Design

• Transmission rate (cont)
• It is possible that the current transmission
  rates for some of these flows are insufficient to
  capture available capacity (unbottlenecked)
• If the reception rate for the last caravan was
  less than the transmission rate, the controller
  attempts a rate decrease
• The newly available capacity is distributed among
  all unbottlenecked flows

29
Design

• Transmission controller
• The job of the transmission controller is simple
  to ensure that each caravan is delivered
  successfully
• Each caravan is streamed (using the rate and
  coding specified by the decongestion controller)
• Until the sender receives an ACK indicating that
  the entire caravan has been successfully received
  and decoded
• The transmission controller can then start
  sending the next caravan

30
Design

• Example
• A and B attempt to send data to C simultaneously
  using Achoo
• Which results in two flows of 10 Mbps each
  arriving at a router R.
• both As flow and Bs flow will achieve an
  end-to-end goodput of about 7 Mbps.

• A decides to start a flow to D
• It divides 10 Mbps between two flows, the B-C
  flow will consume the remaining 9 Mbps on the R-C
  link
• A-D flow is bottlenecked (capacity 2 Mbps, A
  sends at 5 Mbps).
• Bandwidth is decreased on the A-D flow until the
  A-C flow becomes bottlenecked again at 7 Mbps.

31
Challenges

• What about coding overhead?Coding increases the
  end-to-end delay, packet overhead, and
  computational cost of decongestion control
• What about the control channel?Connection
  establishment and teardown for Achoo is the same
  as that for TCPThe challenge is to ensure that
  control messages are not lost in the fray of
  competing data packets
• What about unconventional routing?In recent
  researchers flows can be redirected along
  different paths as traffic conditions change, and
  packets can be sent across multiple paths
  simultaneouslySo coding and transmission rates
  need to be discovered after route changes

32
References

• Barath Raghavan and Alex C. Snoeren. Decongestion
  Control. University of California, San Diego
• P. Maymounkov. Online codes. Technical Report
  TR2002-833, New York University, 2002.
• A. Akella, R. Karp, C. Papadimitrou, S. Seshan,
  and S. Shenker. Selfish behavior and stability of
  the internet A game-theoretic analysis of TCP.
  In Proceedings of ACM SIGCOMM, 2002.
• R. Pan, B. Prabhakar, and K. Psounis. CHOKe - A
  stateless queue management scheme for
  approximating fair bandwidth allocation. In
  Proceedings of IEEE INFOCOM, 2000.
• James F.Kurose and Keith W. Ross Computer
  networking

33
Thanks!