Optimal Internal Congestion Control in A Cluster-based Routernetworks/gswnr2009/Qinghua... ·...
Transcript of Optimal Internal Congestion Control in A Cluster-based Routernetworks/gswnr2009/Qinghua... ·...
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
Optimal Internal Congestion Control in ACluster-based Router
Qinghua Ye
Nov.17, 2009
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
Congestion in the Cluster-based Router
Optimal Utility-based Control
Simulation With NS-3
Evaluations In the Real System
Conclusions and Related Work
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
Figure: Cluster-based Router Architecture
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
Figure: IP Forwarding Path
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
Optimal Utility-based Control
I An optimization approach to congestion control problemsI Objective: maximize the aggregate source utilityI Constraints: network link capacities.
I The network links and traffic sources are viewed as adistributed system that acts to solve the optimization problem
I Traffic sources adjust their transmission rates in order tomaximize their own benefit
I The network links adjust bandwidth prices to coordinate thesources decisions on the evolution of their transmission rates
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
Classification of Optimal Utility-based Control
According to the controlled objects:
I Primal algorithms (TCP)
I Dual algorithms (Active Queue Management)
I Primal-dual algorithms (Combination of TCP and AQM)
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
Internal Congestion Control As An Optimization Problem
I Consider a network with unidirectional links. There is a finiteforwarding capacity C associated with the egress. The egressis shared by a set S of sources, where source s ∈ S ischaracterized by a utility function Us(xs) that is concaveincreasing in its transmission rate xs to the egress.
I Model:P :
∑s∈S
Us(xs) (1)
subject to ∑s∈S
xs ≤ C (2)
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
Decentralized Approach
I The dual theory of optimization leads us to a distributed anddecentralized solution which results in the coordination of allsources implicitly
I Lagrangian function:
L(x , p) =∑s∈S
Us(xs)− p(∑s∈S
xs − C )
=∑s∈S
Us(xs)−∑s∈S
xs ∗ p + p ∗ C(3)
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
Decentralized Approach
I The objective function of the dual problem:
D(p) = maxxs
L(x , p)
=∑s∈S
max(Us(xs)− xs ∗ p) + p ∗ C (4)
I The dual problem:D : min
p≥0D(p) (5)
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
Decentralized Approach
I The congestion control problem can be generalized to tasks offinding distributed algorithms that can make sources adapttransmission rates with respect to the egress price and makeegress adapt prices with respect to loads
I The optimal solution to the distributed congestion controlproblem satisfies:
{∂D(p)∂xs
= ∂Us(xs)∂xs
= U ′s(xs)− p = 0∂D(p)∂p =
∑s∈S (−xs) + C = 0
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
Discrete Optimal Utility-based Control
I To reduce the overhead of transferring the link price, we onlysend the price from the egress to the sources at the beginningof each control interval, which results in a discrete-timecontrol model:
{xs(k + 1) = [xs(k) + K ∗ xs(k) ∗ (U ′s(xs(k))− p(k))]+xs [k]
= [xs(k) + K ∗ (W − xs(k) ∗ p(k))]+xs [k]p(k + 1) = [p(k) + (
∑s∈S xs(k)− C )/R]+p(k)
(6)Here
[g(x)]+y = { g(x), y > 0max(g(x), 0), y = 0
and K and 1/R are step sizes.
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
Queue Status as an Indicator of Congestion
I In real system, the transmission capacity of the egress in themodel vary for different situations or times
I More than one port may share the same busI Sharing of a single egress port by multiple egress queues
I Queue-based approach:
{xs(k + 1) = [xs(k) + K ∗ (W − xs(k) ∗ p(k))]+xs [k]p(k + 1) = [p(k) + (delta(q))/R]+p(k)
(7)
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
Queue Status as an Indicator of Congestion
I The system may be stable at large queue length
I To reduce the stable queue length:
{xs(k + 1) = [xs(k) + K ∗ (W − xs(k) ∗ p(k))]+xs [k]p(k + 1) = [p(k) + (delta(q) + f (q))/R]+p(k)
(8)
I Let f (q) = (q − qo) ∗ u, where qo is the objective of egressqueue length and u is the degree that the queue length wouldbe taken into the price calculation.
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
IP Packet BECN
Adjust Scheduler Parameters
Receive
IP Header Check
Internal Transmit
External
IP Lookup
Check Queue Status
and Generate BECN
Get MAC of Internal Network Device
Packet Classifier
...
Packet Scheduler
Get Mac of External Network Device
Internal
Packet Classifier
To External
To Internal
Local
Forward To Up Layer
External Transmit
External Transmit
Figure: IP Forwarding Path in SimulationQinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
0
500000
1e+06
1.5e+06
2e+06
2.5e+06
0 100 200 300 400 500
Tra
nsm
issi
on R
ate
Time
Transmission Rate Behavior - (K:100000, R:500000000000)
Transmission RateReception Rate
Reception Rate from Ingress 1Reception Rate from Ingress
Reception Rate from Ingress 3
Figure: Optimization utility-based scheme transmission rate behavior
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
0
500000
1e+06
1.5e+06
2e+06
2.5e+06
0 50 100 150 200
Tra
nsm
issi
on R
ate
Time
Transmission Rate Behavior - (W:50000, Q:100)
Transmission RateReception Rate
Reception Rate from Ingress 1Reception Rate from Ingress 2Reception Rate from Ingress 3
Figure: AIMD scheme transmission rate behavior
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
0
100
200
300
400
500
600
700
800
900
1000
0 100 200 300 400 500
Que
ue L
engt
h
Time
Queue Length - (K:100000, R:500000000000)
Queue Length
Figure: Optimization utility-based scheme queue behavior
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
0
200
400
600
800
1000
0 50 100 150 200
Que
ue L
engt
h
Time
Queue Length - (W:50000, Q:100)
Queue Length
Figure: AIMD scheme queue behavior
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
0
200000
400000
600000
800000
1e+06
1.2e+06
0 100 200 300 400 500
Tra
nsm
issi
on R
ate
Time
Transmission Rate Behavior - (K:100000, R:500000000000)
Transmission RateReception Rate
Reception Rate from Ingress 1Reception Rate from Ingress
Reception Rate from Ingress 3
Figure: Fairness - Optimization utility-based scheme transmission ratebehavior
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
0
1000
2000
3000
4000
5000
10 20 30 40 50 60 70 80 90 100
Pack
et R
ate
- K
Pac
ket P
er S
econ
d
Input Rate at Ingress Nodes(Percentage of Wire Rate)
Optimal Utility-based VS. AIMD VS. Original
Reception Rates at Ingress Nodes - originalInjection Rates at Ingress Nodes - original
Transmission Rate at Egress Node - originalReception Rates at Ingress Nodes - AIMD
Injection Rates at Ingress Nodes - AIMDTransmission Rate at Egress Node - AIMD
Reception Rates at Ingress Nodes - optimalInjection Rates at Ingress Nodes - optimal
Transmission Rate at Egress Node - optimal
Figure: Transmission rate comparison
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
-200
0
200
400
600
800
1000
1200
0 20 40 60 80 100
Out
put Q
ueue
Len
gth
Input Rate at Ingress Nodes (Percentage of Wire Rate)
Queue Length With Increasing Offered Traffic
output queue length - originaloutput queue length - AIMD
output queue length - optimal
Figure: Queue variance comparison
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
0
200
400
600
800
1000
298298298
AIMD
298298298
Optimal
298298224
AIMD
298298224
Optimal
298298149
AIMD
298298149
Optimal
29829875
AIMD
29829875
Optimal
Pack
et R
ate
- K
Pac
kets
Per
Sec
ond
Offered Rate at Ingress Nodes(K Packets Per Second)
Fairness Among Different Ingress Nodes
Reception Rate at Ingress Node 1Reception Rate at Ingress Node 2Reception Rate at Ingress Node 3
Injection Rate at Ingress Node 1Injection Rate at Ingress Node 2Injection Rate at Ingress Node 3
Transmission Rate at Egress Node
Figure: Fairness comparison
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router
OutlineCongestion in the Cluster-based Router
Optimal Utility-based ControlSimulation With NS-3
Evaluations In the Real SystemConclusions and Related Work
I Optimal Utility-based Congestion ControlI Fair to different flowsI Efficient to reduce the injection rates of traffic to the internal
network to avoid congestionI Related Work
I Analyze and improve the Internet congestion control schemessuch as TCP and AQM
I In wireless cross-layer congestion control:I Lijun Chen , Stevenh. Low , Mung Chiang , John C. Doyle,
”Optimal cross-layer congestion control, routing andscheduling design in ad hoc wireless networks”
I WeiQiang Xu, etc., ”Dual decomposition method for optimaland fair congestion control in Ad Hoc networks: Algorithm,implementation and evaluation”
I Matthew Andrews, ”Joint Optimization of Scheduling andCongestion Control in Communication Networks”
I Danhua Zhang, Chao Zhang and Jianhua Lu, ”Jointcongestion control, contention control and resource allocationin wireless networks”
Qinghua Ye Optimal Internal Congestion Control in A Cluster-based Router