Loss Discrimination

1
Loss Discrimination

• Vijay Subramanian

2
Background

• It is difficult to determine the cause of packet
  losses at the TCP sender.
• Losses can be due to
• Congestion
• Transmission Errors
• Loss Discrimination/Differentiation Process of
  figuring out the cause of a lost packet.
• Loss Notification Process of informing the
  sender of the cause of loss.

3
Classes of Loss Discrimination

• Implicit Loss Discrimination
• Sender infers cause of packet loss without any
  specific mechanisms.
• Explicit Loss Discrimination
• Sender uses explicit mechanisms to determine the
  cause of a packet loss.

4
Requirements

• In wireless scenarios, transmission losses may be
  mistaken for congestion losses.
• Sender may reduce the sending rate needlessly.
• If sender has the ability to distinguish between
  wireless and transmission losses, it may be able
  to send at an appropriate rate.

5
Requirements

• Sender has to obtain feedback about losses in the
  network.
• Feedback can be provided by
• Network
• With the help of a receiver.
• Without any help from the receiver.
• Receiver
• Cause of loss is inferred by the sender without
  any help.

6
General Framework

• We consider the following scenario.
• Multiple wireless hops on the path from the
  sender to the receiver.
• Intermediate routers are capable of packet
  marking. For example, ECN.

7
End-to-End Loss Probability Scheme

• Requires assistance from the network.
• Augment the IP header with a new field called PLP
  (Path Loss Probability).
• This PLP field is updated at each router along
  the path from the source to destination.
• The receiver echoes the PLP field in the ACK
  packets.

8
What is PLP?

• The PLP field gives the TCP sender an estimate of
  the probability of a packet being dropped due to
  a transmission error on the way from the source
  to the destination.
• This need not be restricted to TCP. If
  implemented in IP, PLP simply gives the
  probability of E2E packet loss probability.

9
What does the network do?

• At each router, the MAC layer keeps a count of
  the number of packets that have been lost due to
  wireless transmission errors as compared to the
  total number of packets sent.
• pf probability of packet loss
• number of packets lost
• number of packets sent in total

10
What does the network do?

• The router can compute the probability of
  success. Ps 1 Pf
• At the TCP sender, the PLP field is set to 1..
• At each router the PLP field, the PLP field is
  multiplied by the local value of the probability
  of success.
• At the receiver , the value of the PLP is echoed
  back to the sender.

11
What does the network do?

• When the sender get values of PLPs in ACK
  packets, it gets the probability of successfully
  transmitting a packet along the path chosen.
• Thus the TCP sender has an idea of the
  probability that a packet sent on multiple
  wireless hops will successfully reach the
  receiver.

12
Advantages

• Use of a PLP field can give the TCP sender an
  accurate estimate of the end-to-end path loss
  probability for a packet since it uses empirical
  data.
• The different routers can update the probability
  of failure of transmission over a particular link
  locally and independently.

13
Disadvantages

• The scheme is not an end-to-end scheme.
• Functionality in routers, TCP sender and receiver
  will have to be changed.
• Routers will have to
• keep track of fraction of packets that were lost
  due to transmission errors.
• Update the value in the PLP field.
• Communication between TCP and IP layers ate
  sender and receiver will be needed.

14
Loss Probability Measurement

• How do the intermediate routers measure the
  packet loss probability on different links?
• Solution
• The MAC protocol used for wireless transmission
  has to measure the number of packets that could
  not be successfully sent over the link.
• For both packet erasures and bit errors, we
  assume that the sender on the link knows whether
  or not packets made it across successfully.
  (Packets are ACKed).

15
Mac Issues

• MAC layer will have to communicate this
  information to IP which updates the PLP field.
• Can we assume that all MAC protocols will receive
  ACKs for their transmissions.
• What does an IP layer do if MAC is unable to
  provide a value for the PLP field? (Nothing?)

16
Questions

• Should the PLP field be in TCP header or IP
  header?
• If in TCP, then routers will have to update TCP
  header field and CP checksum, which is a bad
  idea.
• If in IP header, then receiver IP layer has to
  forward the value in PLP to the TCP receiver. How
  does the TCP receiver then add this value to IP
  header of ACK packets?
• How does ECN do this?

17
ECN Scheme (version 1)

• In this scheme, we use Explicit Congestion
  Notification (ECN) bit to discriminate between
  congestion and transmission losses.
• Assumptions
• TCP sender, receiver and network are ECN capable.

18
ECN Scheme (version 1)

• There is no change to intermediate routers.
• Routers simply set the ECN bit when congestion is
  imminent.
• Receivers simply echo ECN bits back to the sender.

19
ECN Scheme (version 1)

• At the sender, the TCP layer behaves as follows
• If ECN was received and a packet is subsequently
  lost, the loss was most probably due to
  congestion.
• If ECN was not received and a packet was lost,
  assume that the packet was lost due to errors.

20
Advantages

• The only change is to the TCP sender, assuming
  that the network is doing ECN and that sender and
  receiver are ECN capable.
• The change is minimal.

21
Disadvantages

• What happens when a packet with ECN is lost due
  to transmission errors?
• If the receiver does nothing except echo ECN bits
  back to the sender, then sender will assume that
  the packet was lost due to congestion. (which is
  OK)
• If FEC is used, receiver may recover the data in
  the packet and sender will not know that there
  was congestion. What can be done to avoid this?

22
ECN Scheme (version 2)

• Alongside the ECN bit, we add another bit called
  Explicit Wireless Loss Notification (EWLN).
• When a router is unable to transmit a packet
  across a link, it aims to inform the TCP sender
  that the packet belongs to that a packet was lost
  due to transmission errors.
• Ideally, the EWLN bit will be set only in a
  subsequent packet that belongs to the same flow.

23
ECN Scheme (version 2)

• However, this incurs the overhead of maintaining
  flow state at routers which is unacceptable.
• So, the router can set the EWLN bit in all
  packets that flow through it irrespective of flow
  for some specified time.
• There is a good chance that another packet that
  belongs to the same flow will get its EWLN bit
  set.

24
ECN Scheme (version 2)

• The receiver will echo the EWLN bit back to the
  TCP sender.
• Scenario 1
• The TCP sender will get a packet with EWLN bit
  set and will note that a packet has been lost.
• It infers that the packet was lost due to
  transmission errors.
• Scenario 2
• The TCP sender does not receive a packet with
  EWLN bit set even though a packet was lost on a
  link.

25
ECN Scheme (version 2)

• In both scenarios, other TCP senders may receive
  packets with EWLN set even though no packet was
  lost.
• So, a TCP sender should check for recent EWLN
  packets only if a packet has been lost.

26
Related Work

• Some schemes that have been proposed include
• TCP Santa Cruz
• Using packet inter-arrival times.
• ETEN
• Checksum based Loss Discrimination
• These schemes are discussed in brief below.

27
Using Inter-arrival times

• Using Inter-arrival times at the receiver
• Receiver has a better view of the losses
• Assumptions
• Only the last hop is wireless.
• Wireless link is the bottleneck of the
  connection.
• Sender performs bulk transfer.
• Uses a heuristic based approach.
• Claims Works well under some specific
  conditions.

28
Explicit Transport Error Notification (ETEN)

• Explicit Loss Notification strategy

29
TCP Santa Cruz

• Identifies the onset of congestion and the
  direction of congestion.
• Monitors changes in the bottleneck queue length
  over an interval equal to the amount of it takes
  to transmit one window of data and receive
  acknowledgements.
• Monitors the queue developing at the bottleneck
  link.
• Does not use RTT samples.

30
Checksum-based Loss Discrimination

• Use checksums already present in the protocol
  stack.
• Checksum errors do not occur due to congestion.
• Assumes that wireless link is the last hop.
• Different modifications may be needed for
  different link-layer protocols.
• Advantage is that only the receiver has to be
  modified.

31
ECN Scheme

• We change behavior of TCP Sender in response to
  ECN.
• No changes to TCP receiver.
• Assumptions
• Both transports are ECN capable.
• Network is capable of supporting ECN.

32
Assumed Scenario

• Assumptions
• Multiple wireless links
• ECN capable transport and network

Simulation setup
33
Proposed Scheme

• TCP uses two mechanisms to detect losses.
• Timeouts
• Duplicate ACKs (3)
• Upon observing a timeout event, TCP behaves as
  normal.
• Upon seeing a 3 duplicate ACKS, instead of
  halving the window automatically, TCP checks
  whether an ECN mark has been received in the last
  RTT.

34
Proposed Scheme

• If an ECN packet has been received in the last
  RTT, it means that the packet loss was most
  likely due to congestion. In this case, the
  window is reduced.
• On the other hand, if the packet loss was not
  preceded by an ECN packet, the loss was most
  likely due to a transmission error. In this case,
  window is not reduced.
• Thus, the default behavior of TCP to packet
  losses is changed. We now cut the window only to
  ECN markings and not to all packet losses.

35
Will this scheme work?

• If a packet marked with ECN is dropped, this
  scheme will conclude that the loss was due to
  transmission error incorrectly.
• If the packet loss rate is small, then the
  probability that a packet with ECN marking will
  be dropped is small.
• If the loss rate is high, the probability that a
  packet marked with ECN is dropped will be higher.

36
ECN and AVQs
Adaptive Virtual Queue used. AVQs are capable of
marking packets with ECN.
37
Soft State

• To mitigate the problems caused by the corrupted
  ECN packets, routers can maintain soft state
  about flows.
• Soft state Short-lived state about flows that
  are used to mark more than one packet belonging
  to one flow.
• The router extracts the values of source ip,
  Destination IP, source port and destination port

38
Modifications to packet marking

• A packet marked with ECN should not be lost
  later.
• To avoid this, we mark multiple packets with ECN.
  
• At the sender, multiple ECN packets within 1
• RTT will be treated as a single ECN packet.
• TCP sender will not respond to packet losses
  unless they are preceded by an ECN within 1 RTT.

39
Header Hashing scheme at router

• To ensure that ECN packets reach the sender,
  routers mark two packets with ECN.
• The probability of losing 2 ECN packets due to
  transmission errors is much less than the
  probability of losing only 1 packet.
• Routers extract the parameters that identify a
  flow and hash it. A bitmap is used to check if an
  ECN bit was set for this flow recently.

40
Header Hashing scheme (continued)

• If the ECN bit was set for this flow recently,
  the bit is set again.
• Effectively, two packets are marked with ECN
  instead of one.
• Use of a bitmap makes this extremely fast.
• Other flows are not affected.

41
How do we mark packets?
42
To Recap..

• Multiple ECN packets in the same window or within
  1 RTT are considered to be reflective of the same
  loss event/congestion.
• Hence, there is no harm in receiving multiple ECN
  packets together.
• Routers can now mark 2 packets with ECN.
• The probability of losing both is very small.
• If a packet is dropped by a buffer, subsequent
  packets from that flow are marked. Nominally, 2
  packets could be marked.

43
Open questions

• How long should the router wait for the second
  packet?
• What if the second ECN packet is delayed and
  arrives more than 1 RTT after the first one?
• Time stamps could be used at routers to indicate
  when the bits were set.

44
Initial Simulations

• To test whether we can predict losses correctly.

45
References

• Johan Garcia, Anna Brunstrom Transport Layer
  Loss Differentiation and Loss Notification
• S. Biaz and N. H. Vaidya, "Discriminating
  Congestion Losses from Wireless Losses using
  Inter-Arrival Times at the Receiver", IEEE
  Symposium ASSET'99,Richardson, TX, USA, March
  1999.
• S. Biaz and N. H. Vaidya, Distinguishing
  Congestion Losses from Wireless Transmission
  Losses'', Seventh International Conference on
  Computer Communications and Networks (IC3N), New
  Orleans, October 1998.
• S. Biaz and N. Vaidya, "Using End-to-end
  Statistics to Distinguish Congestion and
  Corruption Losses A Negative Result," Texas AM
  University, Technical Report 97-009, August 18,
  1997.
• Rajesh Krishnan ,James P.G. Sterbenz ,Wesley M.
  Eddy , Craig Partridge , Mark Allman Explicit
  Transport Error Notification (ETEN) for
  Error-Prone Wireless and Satellite Networks
• "Decentralized Adaptive ECN Algorithms" published
  in the Proceedings of Infocom2000, Tel-Aviv,
  Israel, March 2000.