TCP Probe: A TCP with Built-in Path Capacity Estimation

1
TCP Probe A TCP with Built-in Path Capacity
Estimation

• Anders Persson, Cesar Marcondes, Ling-Jyh Chen,
  Li Lao,
• M. Y. Sanadidi, Mario Gerla
• Computer Science Department
• University of California, Los Angeles

2
Motivation

• Knowledge of narrow link capacity is important
  for
• Optimizing network utilization via better
  congestion control adaptive streaming
• Tracking dynamic changes in capacity due to
  Vertical Handoffs

3
Bandwidth Estimation Techniques

• Active
• Send out-of-band packets into the network
• CapProbe SIGCOMM04 and Pathrate INFOCOM01
• Passive
• Use ongoing data packets without additional
  overhead to the network
• TCP Probe similar to CapProbe technique but
  uses only actual data packets

4
CapProbe Recap

• Packet pair technique
• Compression expansion due to cross traffic
  distorts Tb
• Filter out distorted samples by using samples
  with minimum end-to-end delay sum
• Fast and accurate estimation of bottleneck
  capacity

5
TCP Probe

• Idea Use CapProbe technique but passively within
  TCP
• Rely on dispersion of ACK pairs and filter
  distorted dispersion based on end-to-end delay
  sum
• Have to ensure that data packet pairs
  corresponding ACK pairs produce accurate capacity
  estimates

6
Challenge

• TCP implementations try to reduce network
  overhead by sending one ACK for two data packets,
  but
• We want an ACK for each data packet!

7
Solution

• Sender sends inverted data packets
• Receiver replies a duplicate ACK and an
  incremental ACK

8
Packet Size Issue

• Probing packet size
• forward direction TCP data 1500 bytes
• reverse direction TCP ACK 40 bytes
• The newly developed AsymProbe Networking05 has
  shown
• If , TCP Probe estimates the capacity
  of the forward direction link
• If , TCP Probe estimates the capacity
  of the reverse direction link

9
Implementation

• Implemented in NS-2 and Linux 2.4 Network Stack
• Approach
• Periodically mark data packet pairs
• Sender timestamps the data packets and the
  corresponding ACKs (?sec granularity)

10
Simulation Scenario

• NS-2 simulation
• Compare TCP Probe capacity estimation with TCPW
  BE (bandwidth estimation)
• TCPW BE is based on dispersion of ACK packets
  without filtering samples

11
TCP Probe vs. TCPW

• 20 flows of the same kind
• TCP Probe All flows estimate C10 Mbps
• TCPW Flows estimate BE between 0 to 7 Mbps

Capacity Estimation (Mbps)
TCP Probe
TCPW BE
12
TCP Probe vs. TCPW

• 1 TCP Probe and 5 TCP New Reno flows
• 1 TCPW and 5 TCP New Reno flows

Capacity Estimation (Mbps)
TCP Probe
TCPW BE
13
Internet Measurement

• Three different paths
• Los Angeles 3 Mbps (cable modem)
• China 45 Mbps
• Alabama 100 Mbps

14
Cable Modem (3 Mbps)
15
China (45 Mbps)
16
Alabama (100 Mbps)
17
Statistical Analysis

• 30 measurements to each destination

18
Applications

• TCP Probe in Vertical Handoff Scenario
• A vertical handoff involves two different network
  interfaces
• Usually represent different technologies and thus
  result in a drastic change in link capacity
• How to get advantage of the extra capacity when
  we are in congestion avoidance?
• Revised TCP Probe
• Re-initiate the estimation every n samples (n50)
• Detect a huge increase in the capacity estimation
• Trigger a new slow start phase (i.e. fast rate
  adaptation)

19
Simulation Scenario

• TCP Probe 1 ? 6
• Pareto flows 7 ? 10, 8 ? 9, 11 ? 14, 12 ? 13
• Capacity between node 1 and 2 increases from
  10Mbps to 100Mbps at 80 second

20
Fast Rate Adaptation
21
Conclusions

• TCP Probe passively provides to a TCP flow sender
  an accurate estimate of path capacity
• Estimated capacity can be used as an upper bound
  on TCP sender rate
• Vertical Handoff applications can benefit from
  detecting capacity change
• Work in progress
• TCP Probe provides improved performance under
  high loss rate (5)
• Astart INFOCOM04 can benefit from accurate
  capacity estimation

22
Adaptive ERE in TCPW

• Sender calculates approximate sending rate ERE
  (Eligible Rate Estimate)
• BE (bandwidth estimation) is aggressive
• RE (rate estimation) is conservative
• RE ERE BE depending on congestion level