Delay-based Congestion
Control for Multipath TCP
Yu Cao, Mingwei Xu, Xiaoming Fu
Tsinghua University
University of Goettingen
Outline

Background and problem statement

Congestion Equality Principle

Weighted Vegas

Simulations

Conclusions
Multipath Transfer

Ever-increasing multihomed hosts

Split traffic across multiple paths.

Provide a new opportunity for designers to
enhance performance of end-to-end transmission.
Benefits for End-hosts


Increase throughput
Improve robustness
[MPTCP, NSDI 2011]
WiFi:
high rate
unstable
low coverage
3G:
low rate
stable
high coverage
Benefits for networks

Bandwidth can be more fairly and
efficiently shared by flows.
S1
S2
S3
6M
9M
D1
D2
D3
New Requirements for MPCC

Improvement on throughput is constrained
by fairness.
S1
S2
S3

9M
D1
D2
9M
9M
D3
Traffic engineering at end-systems
Coupling Subflows Together

Regard network resources as a whole to
compete for bandwidth
S1
S2
S3
6M
9M
D1
D2
D3
S1
S2
D1
15M
D2
S3
D3
How to determine appropriate rates on each path?
How to shift traffic with only local knowledge for sources ?
Congestion Equality Principle

A fair and efficient traffic shifting implies that
every flow strives to equalize the degree of
congestion that it perceives on all its available paths.
A knob to control rates
A metric to estimate congestion degree
Delay-based vs. Loss-based
Packet queuing delay
Packet loss events
Multi-bit info quantifing
congestion degree
Single-bit congestion signals
Be sensitive to …
Perceive changes of congestion
in a large timescale
RTT fairness
Bias against large-RTT flows
Low buffer consumption
Frequent losses
Less aggressively
More aggressively
--
Linked Increases, CMT/RP
Understanding TCP-Vegas
cwnd
rtt
baseRTT
cwnd 
 cwnd
diff  

  baseRTT
rtt 
 baseRTT
  , cwnd   1

diff  
  , cwnd   1

cwnd
x
rtt
diff  
q  rtt  baseRTT
The number of
backlogged packets
x

q
Bandwidth Sharing
 3
 2
 1
3M
6M
2M
1M
Weighted Vegas
Core algo.: allocate alpha to each subflow.
6M
5M
15M
 5
 5
 5
5M
5M
1:4
9M
1M
4M
5M
5M
1:4 ?
5 1 5  4

6
9
To equalize congestion degree of the two paths.
5 5

6 9
Network Utility Maximization
max U s ( ys )
min  Ls (λ )  λcT
s.t. y  Bx


Ls (λ ) : max U s   xs , r    qr xs , r
xs ,r  0,
rRs   rRs  rRs
x0
s S
Ax  c
λ 0
sS
s
Given a fixed budget, invest it in the cheapest
paths to maximize the utility.
Lowest
queuing
delay
Iteratively Tweaking Weights

The total amount of backlogged packets is fixed
at  s , regardless of the number of subflows.
 s , r (t )  ks , r (t  1) s
Control
rates
xs , r (t ) 
 s , r (t )
qr
xs , r (t )  xs , r (t  1)  1
Update
parameters
ks , r (t ) 
xs , r (t )
x
iRs
s ,i
(t )
Tweak
weights
A summary of weighted Vegas

Runs in the same way as TCP-Vegas on each path.

s

Uses equilibrium rates of subflows to adjust weights.
is allocated to subflows according to weights.
A larger  s means more packets are backlogged in
link queues.
A quite small  s makes wVegas over sensitive to
the noise of RTT.
Simulations

We implemented wVegas and Linked Increases in NS-3.

Focus on the fairness and efficiency

Expect wVegas achieves a fine-grained traffic shifting.
s  10
Iteratively adjust rate
Two bottleneck links
Transmission rate
wVegas
Linked Increases
Fairness on Bottleneck Links
wVegas
Linked Increases
Dynamics of Traffic Shifting
The Domino Effect
Rate complementation between subflows
Conclusions

The Congestion Equality Principle

wVegas can achieve fine-grained traffic shifting.

wVegas relies on the accurate measurement of RTTs.

wVegas and Linked Increases have their own
respective advantages and defects.
 Combine they two together?
Thanks