Congestion Control in the Internet with Mixed Traffic Sources and Heterogeneous Access Networks

1
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Candidacy Yi Pan (ypan_at_ics.uci.edu)
• Committee Members
• Tatsuya Suda (Chair)
• Wei K. Tsai
• Magda El Zarki

2
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Introduction
• Solutions overview
• Wavelet de-noising scheme in the core network
• Smooth handoff scheme at the edge
• Conclusion

3
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• The Future Internet
• Multi-service network
• Carries traffic from more and more applications
• Data transmission FTP, HTTP
• Streaming applications VoIP, RTP/RTCP streams
• Ubiquitous access network
• Integrate many different wireless access networks
  to the Internet
• Different wireless access networks WiFi, GPRS,
  WiMAX

4
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Challenges for congestion control
• In the core
• Application traffic can be feedback controlled
  (responsive traffic) or not (non-responsive
  traffic)
• Bursty non-responsive traffic causes high queue
  fluctuation and bursts of packet losses
• Most responsive traffic is long-term TCP traffic
• At the edge
• Different wireless access networks have different
  available bandwidth
• Transition of an active application session
  between different wireless networks often causes
  transmission rate disruption

5
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks
Non-responsive Traffic Sources
AQM router
Packet loss
Responsive Traffic
Different bandwidth
6
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Introduction
• Solutions overview
• Wavelet de-noising scheme in the core network
• Smooth handoff scheme at the edge
• Conclusion

7
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Solutions
• Wavelet De-noising Scheme for AQM routers in the
  core network
• Classify different traffic patterns and reduce
  the queue fluctuation caused by non-responsive
  traffic
• Smooth handoff scheme at the end-systems
• Smooth transition of data paths as well as the
  transmission rates between heterogeneous wireless
  access networks

8
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Traffic classification for AQM routers in the
  core network
• Existing methods
• Classify and identify different traffic patterns
  through packet header examinations
• Multiple layers of protocol headers in the
  packets are examined to identify different
  application flows
• Apply different policies on different traffic
  flows identified
• Packets identified as bursty non-responsive
  traffic are often dropped

9
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Limitations
• Deep packet inspection may be prohibited by
• Network management policies and security
  protocols (IPSec in VPNs)
• Flow identification is not scalable to large
  number of traffic flows
• Per-flow application traffic patterns may not be
  indicated by protocol headers
• E.g. Rate controlled UDP flows are also
  responsive traffic, while HTTP short transactions
  on TCP are bursty
• Increasing cost in packet header examination with
  increasing link speeds
• Need complex hardware in the data path of a
  router to perform line-speed packet header
  examination
• Dropping packets in bursty traffic may impair the
  performance of important applications such as
  HTTP web browsing

10
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• New features of the proposed wavelet de-noising
  scheme
• Classification of responsive and non-responsive
  traffic based on periodic samples of incoming
  traffic volume
• By analyzing transmission rate reduction of
  incoming traffic in response to AQM packet loss
• No packet header examination
• Non-responsive traffic is filtered and bypass the
  AQM queue
• Filtering a virtual AQM queue length is
  maintained only corresponding to responsive
  traffic volume
• Bypassing available buffer in excess to the
  virtual AQM queue length is used to allow
  non-responsive traffic to pass through the router

11
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Contributions
• No header examination for each packet
• Avoid restrictions on packet header examination
  by network management and security protocols
• Scalable to number of traffic flows in the
  network
• No flow identification required
• Not relying on protocol headers to identify
  application traffic patterns
• Scalable to increasing link speed
• No data plane hardware for complex packet
  processing required
• Scheme implemented at control plane of the router
• No additional packet losses when extra buffer is
  available

12
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Solutions
• Wavelet De-noising Scheme for AQM routers in the
  core network
• Classify different traffic patterns and reduce
  the queue fluctuation caused by non-responsive
  traffic
• Smooth handoff scheme at the end-systems
• Smooth transition of data paths as well as the
  transmission rates between heterogeneous wireless
  access networks

13
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Handoff schemes in the wireless access networks
• Existing handoff schemes
• Only transfer data paths from one access point to
  another
• Do not consider bandwidth differences in
  different wireless access networks
• Issue
• Handoff between two different wireless access
  networks can cause disruption in transmission
  rate
• Transmission rate on the old data path may be too
  high/low on the new data path
• Switching to a new data path may require stop the
  current transmission rate and re-establish a new
  transmission rate
• Results in packet loss or low throughput in
  applications
• Example application suffering from the rate
  disruption Real-time streaming applications

14
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• New features of the proposed smooth handoff
  scheme
• Use multiple paths to reach a single mobile node
• Assign different mobile IP addresses (COAs) to
  different paths reaching a single mobile node
• Exploit different amounts of bandwidth on
  multiple paths to a single mobile node
• To reduce or prevent a packet loss due to hand
  off
• To increase throughput for the mobile node

15
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Contributions
• Allow smooth transition of both data paths and
  the transmission rates during handoff
• Using multiple paths during handoff avoids delay
  in switching data paths
• Establish transmission rates on multiple paths
  during the handoff avoids transmission rate
  disruption during handoff
• Allow optimal usage of available bandwidths on
  multiple paths
• Different available bandwidth on multiple paths
  are used to optimize the video quality during
  handoff

16
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Introduction
• Solutions overview
• Wavelet de-noising scheme in the core network
• Smooth handoff scheme at the edge
• Conclusion

17
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• The proposed de-noising scheme

Traffic entering the AQM virtual queue
18
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Wavelet De-noising scheme overview
• Traffic estimator
• De-noising filter

Estimated noise
19
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Wavelet De-noising scheme overview
• Traffic estimator
• Estimate the periodic cycle of responsive traffic
• Most responsive traffic follows TCP AIMD
  mechanism, resulting in periodic rate changes
• Estimate the AQM packet loss rate corresponding
  to responsive traffic
• De-noising filter
• Removing non-responsive traffic bursts
• Non-responsive traffic bursts occurs in time
  scales different from the dominating periodic
  cycle of responsive traffic

20
Traffic estimator

• Estimate the periodic cycle of responsive traffic

Periodic cycle of responsive traffic
Time shift dtk in traffic traces
Strong correlation Incoming traffic comes back
to the peak rate in dtk after a peak in AQM
packet loss rate
Shifted Incoming Traffic u(tdtk)
AQM packet loss rate p(t)
21
Traffic estimator

• Algorithm
• cross-correlations between the incoming traffic
  and AQM packet loss rate are calculated
• The cycle lengths are decided by the strongest
  cross-correlation

Time shift length
Packet loss rate
Incoming traffic
22
TCP traffic estimator

• Estimate AQM packet loss rate for responsive
  traffic

Linear increasing speed
Incoming TCP Traffic
Nloss Number of packet loss in Tqbusy
Average Packet Loss pavg
Queue busy cycle Tqbusy
AQM packet loss rate
23
Traffic estimator

• The average queue busy cycle length is then
• That gives us the conclusion (Theorem 2.4.1)
• The AQM packet loss corresponding to responsive
  traffic is

Observed packet loss in queue busy period
Queue busy ratio 1
Observed queue busy period
Link capacity
24
Traffic estimator

• That gives us the conclusion (Theorem 3.1)
• The AQM packet loss that allows maximum link
  utilization for the long-term TCP traffic is
• Assuming queue busy ratio is r 1
• The average queue busy cycle length is
• Let target Nloss
• Ttarget
• Dividing Tqbusy by Ttarget, we get the estimation
  on AQM packet loss from the current observation
  pavg

Observed packet loss in queue busy period
Queue busy ratio 1
Observed queue busy period
Link capacity
Busy time of the queue
The target queue busy cycle
25
Traffic estimator

• Simulation results with HTTP-type noise

The slope is the estimated ploss
Our estimation on ploss is stable under HTTP-type
noise (i.e. close to the line corresponding to
the TCP configuration in the simulation)
26
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Wavelet De-noising scheme overview
• Traffic estimator
• Estimate the periodic cycle of responsive traffic
• Estimate the AQM packet loss rate corresponding
  to responsive traffic
• Wavelet De-noising filter
• Removing non-responsive traffic bursts
• Non-responsive traffic bursts occurs in time
  scales different from the dominating periodic
  cycle of responsive traffic

27
Wavelet De-noising Filter

• Applying a threshold-based wavelet filter
• Incoming traffic volume u(t) is decomposed into
  wavelet format

Wavelet decomposition
u(t)
De-noised traffic in which non-responsive traffic
changes are removed
Thresholds at different time scales
High bursts to remove
28
Wavelet De-noising filter

• Threshold vector is the solution to the
  following two problems
• MSE problem

AQM packet loss rate at sample period k
De-nosing error
Loss rate of De-noised Traffic
Loss rate of estimated responsive traffic
Queue size used by responsive traffic
Cumulative volume of de-noised traffic
Extra buffer
29
Wavelet De-noising filter

• Linear approximation of MSE

Linear approximation condition
30
Wavelet De-noisingPerformance evaluation

• Single bottleneck scenario configure i)

50 long-term TCP flows with the same RTT
R0
Router with RED output queue
31
Wavelet De-noisingPerformance evaluation

• Performance metrics in configuration i)

Varying available buffer sizes
The proposed scheme achieves a comparable TCP
goodput rate to BLUE with much less packet loss
rate
32
Wavelet De-noisingPerformance evaluation

• Multiple-bottleneck scenario configuration iii)

25 long-term flows with uniform distributed RTTs
25 long-term TCP flows 10 flows with 60ms RTT
and 15 with 180ms RTT
R1
Router with RED output queue
33
Wavelet De-noisingPerformance evaluation

• Performance metrics in configuration iii)

Two bottlenecks link1 and link2
Bottleneck moves to link2
The proposed scheme achieves the good TCP goodput
in multiple bottleneck scenario with two groups
of TCP flows with different RTTs
34
Wavelet De-noisingPerformance evaluation

• Performance metrics in configuration iii)

Two bottlenecks link1 and link2
Bottleneck moves to link2
The proposed scheme achieves lower packet loss
rate in multiple bottleneck scenario with two
groups of TCP flows with different RTTs
35
Wavelet De-noisingPerformance evaluation

• Computation cost with higher link-speed

Computational cost of the proposed scheme remains
relatively constant with increasing link speed
36
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Introduction
• Solutions overview
• Wavelet de-noising scheme in the core network
• Smooth handoff scheme at the edge
• Conclusion

37
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Basic ideas
• Sending a packet on multiple paths during handoff
  reduces loss
• When a packet is lost on one path due to handoff,
  the packet is still available on the other paths

COA1 is registered to Home Agent and
Corresponding Node and Path1 is used to send
packets to COA1
38
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Basic ideas
• Sending a packet on multiple paths during handoff
  reduces loss
• When a packet is lost on one path due to handoff,
  the packet is still available on the other path

Path2 to COA2 and path1 to COA1 are both used to
multicast data packets to the mobile node
39
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Basic ideas
• Adapt to different amounts of bandwidth
• Perform rate control on multiple paths during the
  handoff

Available bandwidth on each path is detected and
proper transmission rate on each path is
established during the handoff
40
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Basic ideas
• Adapt to different amounts of bandwidth
• Multi layer video transmission on multiple paths
  during handoff

Optimize the usage of different bandwidths for
improved video quality
41
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Multi-path transport protocol design

42
Simulation Settings

• Simulation Scenario

Different Average background traffic volume in
different base stations are explored in simulation
Corresponding Node (source of video traffic)
43
Simulation Settings

• Compared handoff schemes
• Single path schemes with single mobile IP
  binding
• No forwarding no local packet forwarding for
  mobile nodes is performed among base stations
• Basic Mobile IP technique
• Forwarding local packet forwarding service is
  enabled by fast mobile IP handoff protocol among
  base stations
• Represent network layer mobility enhancement
  techniques that repair the packet loss on a
  broken path for an active session

44
Simulation Results

• Results and observations
• Video throughput when the mobile node moves from
  high bandwidth cell to low bandwidth cell

45
Simulation Results

• Reduced packet loss

Multi-path handoff scheme keeps the packet loss
ratio low. Base layer is protected with
near-to-zero loss ratio
With different available bandwidth in the new cell
46
Simulation Results

• Improved goodput
• With protection of base layer, the goodput is
  improved in terms of smooth video frame rate

47
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Introduction
• Solutions overview
• Wavelet de-noising scheme in the core network
• Smooth handoff scheme at the edge
• Conclusion

48
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Conclusion
• Designed a de-nosing framework to improve the
  congestion control in the core network with mixed
  traffic patterns
• No packet header examination
• Avoid network administrative and security
  restrictions on packet header examination
• Do not rely on protocol header to identify
  application traffic pattern
• Scalable to the number of traffic flows
• No flow identification required
• Scalable to increasing line-speed
• Algorithms run outside the data path in the
  routers
• Improved performance in multi-bottleneck
  heterogeneous traffic conditions
• High throughput and link utilization
• Reduced packet loss rate

49
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Designed an end-system smooth handoff scheme
• Allow adaptive congestion control when traversing
  different wireless access networks
• Reduced packet losses during the handoff
• No transmission rate disruption
• Optimize the usage of different available
  bandwidths in different wireless networks for
  smooth transmission of video streaming
• Allow smooth transition of video quality between
  different networks
• Improved video quality during handoff

50
Congestion Control in the Internet with Mixed
Traffic Sources and Heterogeneous Access Networks

• Publication list
• Journal and Book Chapters
• J. Lu, Y. Pan, K. Fujii, et al, Adaptive
  Networks, book chapter in Cognitive Networks
  by Wiley Inc., Sept, 2007
• Y. Pan, M. Lee, J.B. Kim, T. Suda, An End-to-End
  Multi-Path Smooth Handoff Scheme for Stream
  Media, in the IEEE Journal of Selected Areas of
  Communications (JSAC), Special-Issue on All-IP
  Wireless Networks, Vol. 22, No. 4, pp. 653-663,
  May 2004
• Conferences and Workshops
• Y. Pan, Detecting and Filtering Non-responsive
  Traffic in AQM Queues without Packet Header
  Examination, in Proc. of ACM SIGMETRICS08
  Student Thesis Panel, June, 2008, Annapolis, MD
• Y. Pan, W. Tsai, and T. Suda, Applying Wavelet
  De-noising to Improve TCP Throughput in AQM
  queues with Existence of Unresponsive Traffic,
  in Proc. of IEEE ICCCN 2007, Honolulu, HI
• Y. Pan, M. Lee, J.B. Kim, T. Suda, An End-to-End
  Multi-Path Smooth Handoff Scheme for Stream
  Media(short version), in the Proceedings of ACM
  WMASH03 Workshop, page 64-74, Sept. 2003, San
  Diego, CA
• Y. Pan, M. Lee, J. B. Kim, T. Suda, Smooth
  Handoff Scheme for Stream Media with Bandwidth
  Disparity in Wireless Cells, in the Proceedings
  of IEEE CCW03 Workshop, page 9-16, Oct. 2003,
  Dana Point, CA
• Technical Reports
• J. Lu, Y. Pan, S. Yamamoto, and T. Suda, Robust
  Data Dissemination for Wireless Sensor Networks
  in Hostile Environments, Technical Report 08-10,
  Bren School of Information and Computer Science,
  University of California at Irvine
• J. Lu, Y. Pan, J. Wang, A. Yahaya, and T. Suda,
  "A Cross-layer Analysis Model for Wireless Sensor
  Network QoS, Technical Report 08-08, Bren School
  of Information and Computer Science, University
  of California at Irvine
• In submission
• Y. Pan, W. Tsai, and T. Suda, Detecting and
  Filtering nonresponsive Traffic in AQM Queues
  using Wavelet De-noising Techniques, submitted
  to IEEE/ACM Transaction on Networks

51
Thanks!