Using Loss Pairs to Discover Network Properties

1
Using Loss Pairs to Discover Network Properties

• Jun Liu, Mark Crovella
• Computer Science Dept.
• Boston University

2
Discovering Network Properties

• What is the state of the network when packets
  are being dropped?
• E.g., What is the buffer occupancy when packets
  are being dropped?
• For a DropTail queue, indicates buffer size
• For an AQM queue, indicates dropping policy.

3
Benefits and Difficulties

• Why is this useful?
• Characterizing existing networks
• Using information gained to adapt applications to
  network status
• Debugging network elements.
• Why is this hard?
• Dropped packets carry no information to the
  endpoint
• Noisy measurements require robust estimation
  methods.

4
The Basic Idea

• A packet traveling close to a dropped packet sees
  similar network state as the dropped one.
• We propose a method called Loss Pairs to
  characterize the dropping function of a network
  element under some assumptions.

5
Loss Pairs

• A non-dropped packet that is close to a dropped
  packet can inform the endpoint about network
  state seen by the dropped packet.
• We define a loss pair as two packets p1 and p2
  such that
• p2 initially follows p1 with time ? between the
  trailing edge of p1 and the leading edge of p2
• Exactly one of p1 and p2 happens to be dropped in
  the network
• p1 and p2 traverse the same sequence of links and
  routers up to the point one is dropped.
• Generally, we consider ? to be 0.

6
Assumptions For Using Loss Pairs in Practice

1. Most of the packet losses and delays happen at
   the bottleneck.
2. The round-trip path and the bottleneck stay
   stable during measurement.
3. To estimate queue state, packet scheduling must
   be FCFS.
4. To convert queue occupancy to bytes, bottleneck
   bandwidth must be known.

7
Network Setting For Evaluation

• We consider a 3-hop sample network setting.
• Capable of varying cross traffic upstream,
  downstream and at the bottleneck.
• Workload heavy tailed ON/OFF TCP sources.
• Goal Characterize the queue between B and C by
  passive measurements taken at the sender.

Upstream Cross Traffic
Bottleneck Cross Traffic
Downstream Cross Traffic
Bottleneck Link
Visable Traffic
8
Varying Buffer Size Of A DropTail Router

• The queue state seen by a dropped packet is a
  linear function of buffer size.

9
Estimating Buffer Size of DropTail Routers

• Tq the most common RTT of Loss Pairs.
• Tp minimum RTT of all non-dropped packets.
• C link bandwidth of the bottleneck link
• Buffer size C (Tq Tp)

10
Filtering Ability of Loss Pairs

• Normal RTT RTT of Loss Pairs

Corresponds to full buffer
Corresponds to empty buffer
11
Estimation Accuracy UnderLight Cross Traffic

• Each crossing path has 1/10th the sources of the
  main path.
• The estimation results are quite good -- all
  assumptions are met.

12
Estimation Accuracy UnderModerate Cross Traffic

• Each crossing path has 50 of the sources of the
  main path.
• Estimation results are still acceptable --
  assumptions are partially met.
• Queue delays in non- bottleneck queues are
  prominent for small buffer sizes.

13
Estimation Accuracy Under Heavy Cross Traffic

• Each crossing path has as many sources as the
  main path.
• Estimation results are poor for small buffers --
  assumptions are violated.
• However, the results are acceptable on large
  buffers.

14
Effect of Cross Traffic

• Heavy cross traffic upstream , downstream or on
  return path affects accuracy.
• Cross traffic at the bottleneck doesnt affect
  accuracy -- only affects the number of samples on
  Loss Pairs.

15
Characterizing Dropping Curves of AQM Routers

• New assumption packet drops are independent.
• Estimation Method loss pairs with RTT
  x
• Dropping ratio -------------------------------
  
• trial pairs with RTT x
• A trial pair with RTT x is a pair of packets
  with at least one packet not having been dropped.

16
Characterizing the Dropping Curve of a RED Router

• Parameters min_threshold9KB, max_threshold18KB,
  Mp0.1
• Measured By Loss Pairs Actual Behavior

17
Characterizing the Dropping Curve of a BLUE Router

• Parameters B500 pkts, IncrDecr0.0025, Holding
  time 0.01 sec.

18
Related Work

• Measurement of bandwidth by Packet Pairs
• Keshav (1991), Bolot (1993), Paxson (1995),
  Carter Crovella (1996, Bprobe)
• Measurement of bandwidth, delay, mean queue
  occupancy, etc. by individual packet
• Jacobson (1997, pathchar), Downey (1999, pchar),
  Mah (1999, clink), Lai and Baker (2000), Harfoush
  et al. (2001)
• Measurement of loss rate
• Duffield et. al. (1999, MINC)
• Harfoush et. al. (2000, MINT)

19
Conclusion

• We propose a new technique for characterizing the
  packet dropping behavior of network elements.
• By simulation, we have shown that this method is
  effective in characterizing the dropping patterns
  of routers.
• This method can be used to characterize the queue
  management scheme being used at the bottleneck
  link.
• For DropTail queues, this method can be used to
  determine router buffer size.