The War Between Mice and Elephants

1
The War Between Mice and Elephants

• By
• Liang Guo Ibrahim Matta
• In Proceedings of ICNP'2001 The 9th IEEE
  International Conference on Network Protocols,
  Riverside, CA, November 2001.
• Presented By
• Eswin Anzueto

2
Outline

• Introduction
• Analyzing Short TCP Flow Performance
• Architecture and Mechanism RIO-PS
• Simulations
• Discussions
• Conclusions and Future work

3
Introduction

• Internet traffic Most (80) of the traffic is
  actually carried by only small number of
  connection (elephants), while the remaining
  large amount of connections are very small in
  size or lifetime (mice)
• In a fair network environment, short
  connections expect relatively fast service that
  long connection. However, some times we can not
  observe such a nice property in the current
  internet.

4
Introduction (cont)

• Factors effecting the performance of mice
• TCP tries to conservatively ramp up its
  transmission rate to the maximum available
  bandwidth. Therefore the sending window is
  initiated at the minimum possible value
  regardless of what is available in the network.

5
Introduction (cont)

• For short connections, since most of the time the
  congestion window is very small, packet loss
  always requires timeout to detect. (not enough
  packets to activate the duplicate ACK mech.)
• ITO (initial timeout) is very conservative. (no
  sampling data is available), Short Connection
  performance is degraded due to large timeout
  period.

6
Introduction (cont)

• In this paper, we propose to give preferential
  treatment to short flows with help from an Active
  Queue Management policy inside the network.
• We also rely on the proposed Diffserv-like
  architecture to classify flows into short and
  long at the edge of the network.

7
Related work

• Crovella et al 2001 16 and Bansal et al 2001
  17 comment that size aware job scheduling helps
  enhance the response time of short jobs without
  hurting the performance of long jobs.
• Seddigh et al 2 shows the negative impact of
  the initial timeout value on the short TCP flow
  latency and propose to reduce the default
  recommended value

8
Outline

• Introduction
• Analyzing Short TCP Flow Performance
• Architecture and Mechanism RIO-PS
• Simulations
• Discussions
• Conclusions and Future work

9
Sensitivity Analysis for Short and Long TCP Flows

• Transmission time of short TCP flows is not very
  sensitive to loss rate when the
• loss rate is relatively small, but it increases
  drastically as loss rate becomes
• larger.

10
Sensitivity Analysis of Transmission Time

• For small-size TCP flows, increasing the loss
  probability can
• lead to increased variability, while for long TCP
  flows, large
• rate reduces the variability of transmission
  times

11
Factors Effecting Variability

• When Loss rate high TCP Congestion control is
  more likely to enter exponential back off phase,
  which can cause significantly high variability in
  transmission time of each individual packet of a
  flow.
• When loss rate low, TCP either in slow start or
  congestion avoidance phase. This dimension of
  variability is more pronounced for long flows.
• Since the first source of variability is on
  individual packets of a flow, the law of large
  numbers indicates that its impact is more
  significant on short flows

12
Sensitivity Analysis of Transmission Time

• We thus conclude that reducing the loss
  probability is more critical to help short TCP
  flows experience less variations in transmission
  (response) time.
• Observe that the C.O.V. of transmission times is
  closely related to the fairness of the
  systemsmaller values imply higher fairness.
• Such interesting behavior motivates us to give
  preferential treatment to short TCP flows.

13
Preferential Treatment to Short TCP flows

• Assumption
• Giving preferential treatment to short TCP flows
  can significantly enhance their transmission
  time, without degrading long flow performance.
• Simulation using NS simulator
• 10 long(10000-packet) TCP-NewReno flows and 10
  short(100-packet) TCP-Newreno flows over 1.25Mbps
  link.
• Queue Management Policy Drop Tail, RED ,RIO
  with preference to short flows.

14
Link Utilization under Drop Tail, RED and RIO-PS
15
Network Goodput Under Different Schemes
16
Preferential Treatment to Short TCP flows (cont)

• In fact, this preferential treatment might even
  enhance the transmission of long flows since they
  operate in a more stable network environment
  (less disturbed by short flows) for longer
  periods.
• In a congested network, reducing the packet drops
  experienced by short flows can significantly
  enhance their response time and fairness among
  them.

17
Outline

• Introduction
• Analyzing Short TCP Flow Performance
• Architecture and Mechanism RIO-PS
• Simulations
• Discussions
• Conclusions and Future work

18
Proposed Architecture
19
Edge Router

• Determines packet coming from long or short flow.
• Accurate flow characterization can be very
  complicated,
• instead we simply maintain a counter (Lt) that
  tracks how many packets have been observed so far
  for a flow. Once Lt exceeds a certain threshold
  we consider the flow to be long.
• Per flow state information are softly maintained
  to detect the termination of flow. The flow hash
  table is updated periodically every Tu time
  units.
• It is configured with SLR (Short to Long ratio).
  It then periodically (every Tc time units)
  performs AIAD control over the threshold to
  achieve the target SLR

20
Core Router

• Gives preferential treatment to short packets.
• RIO (Red In and Out) queuing policy is used
  because its conformity to the Diff-serv
  specification.
• The probability of dropping short packets depends
  on the average backlog of short packets queue.
  On the contrary, for long packets the total
  average queue size is used to detect incipient
  congestion and those flows have to give up some
  resources.
• No packet reordering will happen in the FIFO
  queue with RIO
• RIO inherits all features of RED

21
RIO Queue with preferential treatment to short
flows
22
Outline

• Introduction
• Analyzing Short TCP Flow Performance
• Architecture and Mechanism RIO-PS
• Simulations
• Discussions
• Conclusions and Future work

23
Simulation
Core router
Edge router
24
Simulation (cont)
25
Simulation (cont)

• 4000 secs simulation time,2000 secs warm up time.
• Average response time relative to RED

26
Simulation (cont)
27
Simulation (cont)
28
Transmission Time of foreground traffic
29
Network goodput
30
Outline

• Introduction
• Analyzing Short TCP Flow Performance
• Architecture and Mechanism RIO-PS
• Simulations
• Discussions
• Conclusions and Future work

31
Discussion Comments on the Simulation Model

• Our simulation is one-way traffic model.
• All TCP connections have similar end to end
  propagation delays, this is not common topology
  seen by internet users
• When there is also reverse traffic present, out
  proposed scheme will even have a superior
  performance over the traditional policies.

32
DiscussionQueue Management Policy

• RIO neither provides absolute aggregate (class
  based) nor relative flow based guarantees.
• To attack these problems, one may resort to AQM
  policies like the PI controlled RED queue , or a
  service model like the Proportional Diffserv
  proposed in for better control over the
  classified traffic and more predictable service

33
DiscussionDeployment Issues

• Our proposed scheme requires edge devices to be
  able to perform per-flow state maintenance and
  per-packet processing.
• The scheme does not require the queue mechanisms
  to be implemented at each router.

34
DiscussionFlow Classification

• We use a threshold based classification method.
• Such method thus mistakenly classifies the first
  few packets of a long flow as if they came from a
  short flow. However, such mistake may help to
  enhance performance and make the system more fair
  to all TCP connections
• The first few packets of a long flow are more
  vulnerable to packet losses and deserve to be
  treated with high preference.

35
DiscussionsController Design

• Our preliminary results indicate that the
  performance is not very sensitive to the target
  load ratio of active short to active long flows
  (the value of SLR at the edge).
• The actual SLR depends on the values of Tc and
  Tu, which determine how often the classification
  threshold and active flow table are updated,
  respectively.

36
DiscussionMalicious Users

• One concern regarding our proposed scheme may be
  that users are then encouraged to break long
  transmissions into small pieces so that they can
  enjoy faster services. However, we argue that
  such initiative may not be so attractive to users
  given the large overhead of fragmentation and
  reassembly.

37
Outline

• Introduction
• Analyzing Short TCP Flow Performance
• Architecture and Mechanism RIO-PS
• Simulations
• Discussions
• Conclusions

38
Conclusions

• The Performance of the majority of TCP flows (the
  short transfers or mice) is improved in terms of
  response time and fairness
• The Performance of few elephants is also improved
• Overall goodput of the system is also improved or
  at least stays almost the same

39
Conclusion (cont)

• The proposed architecture is flexible in that the
  functionality that defines this scheme can be
  largely tuned at the edge routers

40
Thank you!

• Thank you to the following persons for the
  pictures used for this presentation
• Matt Hartling Sumit Kumbhar
• Iris Su
• Preeti Phadnis