The War Between Mice and Elephants

1
The War Between Mice and Elephants

• By
• Liang Guo (Graduate Student)
• Ibrahim Matta (Professor)
• Boston University
• ICNP2001
• Presented By
• Preeti Phadnis

2
Outline

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

3
Mice vs Elephants

• Mice Short TCP flows e.g. Web Traffic
• 20 of internet traffic is carried by large
  number of mice

• Elephants Long TCP flows e.g. FTP
• 80 of internet traffic is carried by small
  number of elephants

4
Internet today

• WWW World Wide Wait term coined by R. Khare
  and I .Jacobs
• Users spend long time downloading a plain text
  webpage
• Reason The mice dont get the fair share of the
  network resources

5
Factors effecting the performance of mice

• TCP tries to conservatively ramp up its
  transmission rate to the maximum available
  bandwidth
• For short connections, since congestion window is
  very small, packet loss always requires timeout
  to detect.
• For the first few packets, since no sampling data
  is available, TCP has to use a conservatively
  estimated ITO value as RTO. Short Connection
  performance is degraded due to large timeout
  period.

6
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.
• D.D Clark and W.Fang 1998 4 AQM scheme using
  RED with In and Out Policy

7
Outline

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

8
Sensitivity Analysis for Short and Long TCP Flows
9
Sensitivity Analysis of Transmission Time
10
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. Short flows are effected more due to this
  reason.
• When loss rate low, TCP either in slow start or
  congestion avoidance phase. This dimension of
  variability is more pronounced for long flows.

11
Preferential Treatment to Short TCP flows

• 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.

12
Link Utilization under Drop Tail, RED and RIO-PS
13
Network Goodput Under Different Schemes
14
Outline

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

15
Proposed Architecture
16
Edge Router

• Determines packet coming from long or short flow
• Maintains a counter Lt that tracks how many
  packets have been observed so far for a flow.
  Lt is dynamic
• 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

17
Core Router

• Gives preferential treatment to mice
• RIO (Red In and Out) queuing policy is used4
  with preferential treatment to short flows-
  RIO-PS
• RIO used twin RED algorithms for dropping packets
  one for ins and one for outs.
• The probability of dropping in packets depends
  on the in average in packet queue and the
  probability of dropping out packets depend on
  the total average queue length.
• No packet reordering will happen in the FIFO
  queue with RIO
• RIO inherits all features of RED
• RIO performs soft prioritization, thus does not
  lose the benefit of statistical multiplexing.

18
RIO Queue with preferential treatment to short
flows
19
Outline

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

20
Simulation Setup
21
Simulation Configuration
22
Experiment 1

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

23
Instantaneous Queue Size and Drop Rate
24
Fairness of Transmission time
25
Transmission Time of foreground traffic
26
Network goodput
27
Experiment 2Unbalanced Requests

• Client set 1 requests smaller objects ,Client set
  2 requests larger objects

28
Experiment 2
29
Outline

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

30
Discussion

• Simulation Model
• Dumbbell and Dancehall model used.
• All TCP connections have similar end to end
  propagation delays, this is not common topology
  seen by internet users
• If reverse traffic present even better
  performance
• Queue Management Policy
• RIO neither provides absolute aggregate (class
  based) nor relative flow based guarantees.
• Other AQM policies like PI controlled RED queue
  better

31
Discussions

• Deployment Issues
• Edge devices need to perform per-flow state
  maintenance and per packet processing but it does
  not effect performance.
• Not required to implement queue policies at each
  router, RIO-PS can be implemented at busy
  bottleneck links.
• Flow Classification
• Threshold based classification classifies the
  first few packets of all flows to be short but it
  helps enhance performance .

32
Discussions

• Controller design
• The actual SLR depends on values of Tc and Tu,
  which determines Lt. Smaller values of these
  increases accuracy at the expense of increased
  overhead
• Malicious users
• Can break long transmissions into short flows but
  overhead of fragmentation and reassembly is very
  high.

33
Outline

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

34
Conclusions and Future Work

• Performance of mice is improved
• Performance of few elephants is also improved
• Overall goodput of the system is also improved
• The proposed architecture is flexible in that the
  functionality that defines this scheme can be
  largely tuned at the edge routers

35
Future work

• Integrate size aware traffic management at both
  network and transport layers