COM 360

1
COM 360
2
Chapter 6

• Congestion Control and
• Resource Allocation

3
Allocating Resources

• How do we effectively allocate resources among a
  collection of competing users?
• These resources include the bandwidth of the
  links and the buffers on the routers, or switches
  where the packets are queued awaiting
  transmission.
• Packets contend at a for the use of a link
  router.
• When too many packets are queued waiting for the
  same link, the queue overflows and packets are
  dropped. When this happens often, the network is
  said to be congested.
• Most networks provide congestion-control
  mechanisms

4
Allocating Resources

• Congestion control and allocating resources are
  two sides of the same coin
• If a network actively allocates resources, such
  as scheduling a virtual circuit, then congestion
  may be avoided.
• Allocating network resources is difficult because
  the resources are distributed throughout the
  network.
• On the other hand, you can send as much data as
  you want and recover from congestion if it
  occurs. This is the easier approach, but it can
  be disruptive.
• Thus congestion control and resource allocation
  involve both hosts and network elements, like
  routers, as well as queuing algorithms.

5
Issues in Resource Allocation

• Resource allocation is complex and is partially
  implemented in routers or switches and partially
  in the transport protocol running on the end
  hosts.
• End systems use signaling protocols to convey
  their resource requirements to network node,
  which reply with information about availability.

6
Terminology

• Resource allocation is the process by which
  network elements try to meet the competing
  demands that applications have for network
  elements.
• Congestion control describes the effort the
  network nodes make to respond to overload
  conditions.
• Flow control involves keeping a fast sender from
  overflowing a slow receiver.
• Congestion control is intended to keep a lot of
  senders from sending too much data into the
  network because of a lack of resources at some
  point.

7
Network Model

• Packet Switched Network
• Problem is the same for routers or switches on a
  network or an internet.
• Source may have sufficient capacity to send a
  packet on its outgoing link, but an intermediate
  link may have heavy traffic.
• For example, 2 high-speed links may feed into a
  low speed link as seen on the next diagram

8
Congestion in a Packet Switched Network
9
Congestion Control

• Congestion control is not the same as routing,
  and routing around a congested link does not
  always solve the problem.
• In the previous example, it is not possible to
  route around the router and this congested router
  is referred to as a bottleneck.

10
Connectionless flows

• In the Internet Model, IP provides a
  connectionless datagram delivery service and TCP
  implements an end-to-end connection abstraction.
• Datagrams are switched independently, but usually
  a stream between a particular pair of hosts flows
  through a particular set of routers.
• The idea of flow- a sequence of packets following
  the same route is an important abstraction in
  connection control.

11
Connectionless flows

• Flows can be defined as host-to-host,
  process-to-process.
• A flow is similar to a channel.
• A flow is visible to routers inside the network,
  and a channel is an end-to-end abstraction.
• A flow can be implicitly defined or explicitly
  established like a connection.

12
Multiple Flows
Multiple flows passing through a set of
routers.
13
Taxonomy

• Resource allocation mechanisms can be
  characterized as
• Router-Centric versus Host-Centric
• Reservation-Based versus Feedback-Based
• Window-Based versus Rate-Based

14
Router-Centric vs. Host-Centric

• In router-centric design, each router takes
  responsibility for deciding when packets are
  forwarded and selecting which packets are dropped
  as well as for informing hosts that are
  generating the network traffic how many packets
  they are allowed to send.
• In host-centric design, the end hosts observe the
  network conditions and adjust their behavior
  accordingly.
• These are not mutually exclusive.

15
Reservation-Based versus Feedback-Based

• Resource allocation mechanisms are sometimes
  classified according to whether they use
  reservations or feedback.
• In a reservation-based system, the end host asks
  the network for a certain capacity at the time a
  flow is established. The router allocates enough
  resources, or rejects the flow.
• In a feedback-based system, the end hosts begin
  sending data and adjust their sending rate
  according to the feedback they receive.

16
Window-Based versus Rate-Based

• Both flow control and resource allocation
  mechanisms need a way to express to the sender,
  how much data they can transmit. They do this
  with a window or a rate.
• In a window-based transport, such as TCP, the
  receiver advertises the window to the sender.
  This limits how much data can be sent a form of
  flow control.
• A rate can also be used to control the senders
  behavior. The receiver says it can process a
  certain number of bits per second and the sender
  adheres to this rate.

17
Evaluation Criteria

• How does a network effectively and fairly
  allocate its resources.
• These are the two criteria by which we can
  evaluate whether a resource allocation mechanism
  is a good one or not.

18
Effective Resource Allocation

• Consider the two principle network metrics
  throughput and delay (latency).
• It may appear that increasing throughput means
  reducing delay, but that is not always the case.
• One way to increase throughput is to allow as
  many packets as possible, driving the utilization
  up to 100.
• But increasing the number of packets, increases
  the length of the queues, which means packets are
  delayed longer in the network.

19
Power of a Network

• The power of the network describes this
  relationship of throughput and delay
• Power Throughput/Delay
• This is based on M/M/1 queues ( 1 server and a
  Markov distribution of packet arrival and
  service).
• This assumes infinite queues, but real networks
  the have finite buffers and occasionally drop
  packets.
• The objective is to maximize this ration, which
  is a function of the load on the network.
• Ideally the resource mechanism operates at the
  peak of this curve.

20
Power Curve
21
Effective Resource Allocation

• Ideally, we want to avoid the throughput going to
  zero because the system is thrashing.
• We want a system that is stable- where packets
  continue to get through the network even when the
  network is operating under a heavy load
• If the mechanism is not stable, the network may
  experience congestion collapse.

22
Fair Resource Allocation

• Fairness presumes that a fair or equal share of
  the bandwidth is allocated to each flow.
• Raj Jain has proposed a metric to quantify the
  fairness of a congestion-control mechanism.
• (See formula p. 461)
• Should we consider the length of the paths being
  compared?
• What is fair when one-four hop flow is compared
  with three one-hop flows?

23
Fairness
One four hop flow competing with three
one-hop flows
24
Queuing Disciplines

• Each router must implement some queuing algorithm
  that governs how packets are buffered while
  waiting to be transmitted.
• The queuing algorithm allocates both bandwidth
  (which packets get transmitted ) and buffer space
  (which packets get discarded).
• It also directly affects delay or latency by
  determining how long a packet waits to be
  transmitted.
• Two common queuing algorithms are FIFO and Fair
  Queuing (FQ)

25
FIFO

• FIFO first in first out the first packet into
  the router is the first to be transmitted.
• Since the amount of buffer space is finite, if a
  packet arrives and the buffer is full, the router
  discards it.
• This is sometimes called a tail drop, since the
  packets that arrive at the tail end of the FIFO
  are dropped.

26
FIFO
a) FIFO queuing b) tail drop at a FIFO
queue
27
FIFO and Priority

• FIFO is the simplest algorithm and is the most
  widely used currently in Internet routers.
• A simple variation is a priority queue. The idea
  is to mark each packet with a priority (in the IP
  (TOS) Type of Service field).
• The routers implement multiple FIFO queues, one
  for each priority class and transmit from the
  highest priority queue first.
• This can cause starvation, when low priority
  packets do not get serviced.
• It is used to give the router updating packets
  highest priority.

28
Fair Queuing (FQ)

• Fair queuing maintains a separate queue for each
  flow currently being handled by the router.
• The router services those queues in a round-robin
  order, giving each a chance in order.
• Since the traffic sources do not know the state
  of the router, this must still be used in
  conjunction with a congestion control mechanism.

29
Fair Queuing (FQ)
A separate queue is maintained for each flow.
30
Fair Queuing Example

• Packets with earlier finishing times are sent
  first
• Sending of packet already in progress is completed

Algorithm selects both packets in a) from flow 1
to be transmitted, because of their earlier
finishing times. In b) the router has begun to
send a packet from flow 2 when, the packet from
flow 1 arrives.
31
TCP Congestion Control

• TCP sends packets into the network without a
  reservation and then reacts to observable events
  that occur.
• TCP assumes FIFO queues, but works with FQ also.
• TCP is said to be self-clocking since it uses the
  ACKs to pace the transmission of packets.
• It also maintains variables such as
  CongestionWindow and MAXwindow and increases and
  decreases the window size.

32
Packets in Transit
Additive increase one packet is added during
each RTT
33
TCP SawTooth Pattern
Typical TCP Sawtooth pattern of continually
increasing and decreasing the window as a
function of time instead of increasing and
decreasing by one as in the additive increase.
34
Slow Start
TCP provides another mechanism used to increase
the congestion window rapidly from a cold start.
It adds one packet then two , etc
35
Behavior of TCP Congestion Control
Blue line is the value of the CongestionWindow
over time Bullets at top are timeouts Hash marks
at top are time when each packet is
transmitted Vertical bars are time when packet
that is retransmitted was first transmitted.
36
Fast Retransmit and Fast Recovery

• Fast retransmit was added to TCP to trigger a
  retransmit sooner that the regular timeout
  mechanism.
• When a data packet is received the receiver sends
  an ACK.When a packet arrives out of order it
  cannot be acknowledged, because the earlier
  packet has not been acknowledged, so TCP sends
  the same ACK it send last time- a duplicate ACK.
• When the sender receives a duplicate ACK, it
  knows that a packet was missing, and retransmits.
• TCP waits for 3 duplicate ACKS before
  retransmitting.

37
Fast Retransmit
Fast retransmit based on duplicate ACKs
38
TCP with Fast Retransmit
Blue line is the value of the CongestionWindow
over time Bullets at top are timeouts Hash marks
at top are time when each packet is
transmitted Vertical bars are time when packet
that is retransmitted was first transmitted.
39
Congestion Avoidance Mechanisms

• TCPs strategy is to control congestion once it
  happens, as opposed to avoid congestion in the
  first place.
• TCP repeatedly increases the load on the network
  to find the point at which congestion occurs,
  then it backs off from this point. (It finds the
  available bandwidth.)
• An alternative is to predict when congestion is
  about to happen and to reduce the rate at which
  hosts send packets, just before the packets start
  being discarded this is congestion avoidance.

40
Congestion Avoidance Mechanisms

• Three different avoidance mechanisms put
  additional functionality into the router to
  anticipate congestion
• DECbit splits responsibility between router and
  end nodes. Router sets a bit if the average queue
  length gt 1 when packet arrives.
• Random Early Detection (RED) each router
  monitors its own queue length and notifies the
  source of congestion.
• Source-based Congestion Avoidance- attempts to
  avoid congestion form the end nodes and watches
  for a sign from the network that some routers
  queue is increasing.

41
Average Queue Length
Computing average queue length at router.
42
Weighted Average Queue Length
43
RED thresholds on a FIFO QUEUE
If average queue length is smaller than lower
threshold, no action is taken. If it is larger
than the upper(MAX) threshold, the packet is
dropped. If it is between the two thresholds then
the packet is dropped with some probability P.
44
Drop Probability function for RED
45
Source-Based
Congestion window vs. observed throughput
rate Top congestion window middle observed
throughput Bottom buffer space taken up at the
router
46
TCP Vegas Congestion Avoidance Mechanism
47
Quality of Service

• Packet Switched networks have promised the
  support for multimedia applications which
  combine, audio, video and data.
• One obstacle to this has been the need for higher
  bandwidth links.
• Improvements in coding and the increasing speed
  of links are bringing this about.

48
Real-time Applications

• Real-time applications are sensitive to the
  timeliness of data delivery- they need assurance
  from the network that the data will arrive on
  time.
• Non-real time applications use retransmission to
  be sure data arrives correctly, but this only
  adds to the delay.
• Timely delivery must be provided by the network
  itself ( the routers) and not just the hosts.

49
Quality of Service

• Applications that are happy with best effort
  service should also be able to use the new
  service model which provides time assurances.
• This implies that the network will treat some
  packets differently.
• A network that can provide different levels of
  service is said to support Quality of Service
  (QoS)

50
Application Requirements

• Divide applications into real-time and non
  real-time or traditional data applications.
• Non real-time applications (like telenet, ftp,
  email, web browsing) are also called elastic
  since they are able to stretch into increased
  delay.
• They do not become unusable with increased delay
  (users just become frustrated!)

51
Real-Time Audio

• In audio, data is generated by collecting samples
  form a microphone and digitizing them using an
  analog-to-digital (A-to-D) converter.
• The digital samples are placed in packets which
  are then transmitted across the network to a
  receiver at the other end.
• At the receiver the data must be played back at
  some appropriate rate, usually the same at which
  they were recorded.
• If data arrives after its playback time, it is
  useless.

52
An Audio Application
53
Real-Time Audio

• One way to make a voice application work would be
  to require all samples to take the same amount of
  time to traverse the network, but it is difficult
  to do this.
• The way to deal with this at the receiver end is
  to buffer some amount of data, providing a store
  of packets waiting to be played back at the right
  time.
• If it has a short delay, it is buffered until its
  playback time arrives if it is delayed a longer
  time may not need to be buffered very long.
• We have effectively added a constant offset to
  the playback time called playback point.

54
A Playback Buffer
55
Delay Distribution on the Internet
97of packets had latency lt 100ms. An average of
3 out of every 100 packets would arrive too late
to be used. The tail of the graph extends to
200ms, which would have to be the playback time
to ensure that all packets arrive on time.
56
Taxonomy of Real-Time Applications

• Characteristics of Real-time applications
• Tolerance for loss of data (some audio can
  sustain a loss of a late packet, but a robot arm
  cannot function if it loses its command to stop)
• Adaptability- some audio may be able to extend
  their playback time to adapt to network delays
  called delay-adaptive
• Rate adaptive can trade off bit-rate verses
  quality ( for some video applications)

57
Taxonomy of Applications
58
Token Buckets
Illustrates how a token bucket can be used to
characterize a flows bandwidth requirements.
Shows two flows with average rates but different
token buckets descriptions.
59
Reservations
Making reservations on a multicast
tree.
60
Assured Forwarding (AF)
RED with in and out drop probabilities
61
Summary

• Resource allocation is both central to networking
  and also a very hard problem.
• Congestion control is concerned with preventing
  degradation of service when the demand for
  services exceeds the available supply of the
  network.
• Different quality of service can be provided to
  applications that need more assurances than those
  provided by the best-effort model.

62
Open Issue

• The larger question we should ask is how much can
  we expect from the network and how much
  responsibility will ultimately fall to the end
  hosts?

63
Figure 6.27
64
Figure 6.28
65
Figure 6.29