optimizing network calculus for switched ethernet network...

Optimizing Network Calculus for SwitchedEthernet Network with DRR Scheduler

Aakash SONI 1,2 Xiaoting LI 2

Jean-Luc SCHARBARG 1 Christian FRABOUL 1

1IRIT Toulouse 2ECE Paris

[email protected]

RTSS - 2018

December 13, 2018

Context : Evolution of Avionic Network

Traditional aircraft network ARINC 429. (Airbus A320)Mono-transmitter buses with limited performances (100Kbits/s).

Avionics Full DupleX (AFDX) network.Switched Ethernet ARINC 664. (Airbus A380)A backbone network for the avionics platform.100 Mbps.FIFO/SP queues.

Aakash SONI Optimizing NC for Avionic Network with DRR 1/42

Context : The Problem

Non intrinsic determinism of resource access in traditionalEthernet.

Collisions on the point-to-point links levelSolution : ARINC 664

Full Duplex Links.No collision frame losses.

ARINC 664 : Indeterminism at Switch level.Competition for the use of the resources.

Congestion = frame lossesFrame storage in queues = Latency andjitter.

Switch

. . .

Input flow traffic

CogestionLatencyJitter

Need of guaranteed bounds for certification.(Motivation)


Context : The Problem

Inefficient use of available bandwidth.[1]

Lightly loaded network (up to 10 % only).

Possibility to share bandwidth among critical(avionic) and non-critical flow.Example:> Audio message from cockpit to cabin.> Parking video.

Solution : Quality of Service (QoS) mechanism.(Motivation)

[1] ] H. Charara, J-L.Scharbarg, J. Ermont and C.Fraboul, ”Methods for bounding end-to-end

delays on an AFDX network,” ECRTS 2006.


Context: Objective 1

How to make better use of available bandwidth?

QoS : Share bandwidth using Round Robin Scheduler.Deficit Round Robin (DRR) scheduling.Weighted Round Robin (WRR) scheduling.[1]

DRR

Priority 1

Priority 2

Priority 3

FIFO BuffersTraffic class

C1

C2

C3

[1] ] A. Soni, X. Li, J-L.Scharbarg, and C.Fraboul, ”WCTT analysis of avionics Switched Ethernet

Network with WRR Scheduling,” RTNS 2018.



Performance analysis of DRR and WRR scheduler in AFDXnetwork.

Worst-case end-to-end delay.

Fairness.

DRR / WRR



Improve delay bound computation.

Reduce pessimism in analysis approach.

Worst-casedelay

Sure upper bound ofworst-case delay

Pessimism

E2EDelay

NC approach

Optimizeddelay bound

Optimization


Table of Contents

1 Context

2 Switched Ethernet Network

3 DRR Algorithm

4 Network Calculus

5 Optimization

6 Evaluation

7 DRR vs WRR

8 Conclusion


Table of Contents

1 Context


3 DRR Algorithm

4 Network Calculus

5 Optimization

6 Evaluation

7 DRR vs WRR

8 Conclusion


Switched Ethernet Network : AFDX Network Model

AFDX network model

End-Systems (ex)

Switches (Sx)

FIFO output ports

Statically defined flows.

e1

e2 e3

e4S1 S2

DRR

DRR

DRR

DRR

DRRDRR

Perspective

AFDX

e1

e2 e3

e4S1 S2

FIFO

FIFO

FIFO

FIFO

FIFOFIFO

ExistingAFDXModel


Switched Ethernet Network : AFDX Flow Model

Avionic flows are characterized as virtual links;

Statically defined : predictable deterministic behavior.

Maximum frame length: Smax

Minimum delay between two consecutive frames: BAG(Bandwidth Allocation Gap)

Multi-cast routing

E1E4S1 S2

v1 v1 v1

S3 E5v3

v1E2

v2

v2v3v4

v1v2v4E3

BAG BAG BAG


Table of Contents

1 Context


3 DRR Algorithm

4 Network Calculus

5 Optimization

6 Evaluation

7 DRR vs WRR

8 Conclusion


DRR Algorithm

DRR

C2C1

Active flow

v4

v3

v2

v1

v7

v6

v5

10

20

30

40

0

Q = 20

∆ : Deficit

Credit = Q+ ∆

QC1 ∆C1 QC2 ∆C2

buffers

bytes


DRR Algorithm

DRR

C2C1

Active flow

v4

v3

v2

v1

v7

v6

v5

10

20

30

40

0

Q = 20

∆ : Deficit

Credit = Q+ ∆

QC1 ∆C1 QC2 ∆C2

buffers

bytes

20


DRR Algorithm

DRR

C2C1

Active flow

v4

v3

v2

v1

v7

v6

v5

10

20

30

40

0

Q = 20

∆ : Deficit

Credit = Q+ ∆

QC1 ∆C1 QC2 ∆C2

buffers

bytes

20

v4v3

v2

v1

0


DRR Algorithm

DRR

C2C1

Active flow

v4

v3

v2

v1

v7

v6

v5

10

20

30

40

0

Q = 20

∆ : Deficit

Credit = Q+ ∆

QC1 ∆C1 QC2 ∆C2

buffers

bytes

20

v4v3

v2

v1

0 20


DRR Algorithm

DRR

C2C1

Active flow

v4

v3

v2

v1

v7

v6

v5

10

20

30

40

0

Q = 20

∆ : Deficit

Credit = Q+ ∆

QC1 ∆C1 QC2 ∆C2

buffers

bytes

20

v4v3

v2

v1

0 20

v7

v6

v5

10


DRR Algorithm

DRR

C2C1

Active flow

v4

v3

v2

v1

v7

v6

v5

10

20

30

40

0

Q = 20

∆ : Deficit

Credit = Q+ ∆

QC1 ∆C1 QC2 ∆C2

buffers

bytes

20

v4v3

v2

v1

0 20

v7

v6

v5

10

20


DRR Algorithm

DRR

C2C1

Active flow

v4

v3

v2

v1

v7

v6

v5

10

20

30

40

0

Q = 20

∆ : Deficit

Credit = Q+ ∆

QC1 ∆C1 QC2 ∆C2

buffers

bytes

20

v4v3

v2

v1

0 20

v7

v6

v5

10

20

v2v1

0


DRR Algorithm

DRR

C2C1

Active flow

v4

v3

v2

v1

v7

v6

v5

10

20

30

40

0

Q = 20

∆ : Deficit

Credit = Q+ ∆

QC1 ∆C1 QC2 ∆C2

buffers

bytes

20

v4v3

v2

v1

0 20

v7

v6

v5

10

20

v2v1

0 20


DRR Algorithm

DRR

C2C1

Active flow

v4

v3

v2

v1

v7

v6

v5

10

20

30

40

0

Q = 20

∆ : Deficit

Credit = Q+ ∆

QC1 ∆C1 QC2 ∆C2

buffers

bytes

20

v4v3

v2

v1

0 20

v7

v6

v5

10

20

v2v1

0 20

v6v5

0


Table of Contents

1 Context


3 DRR Algorithm

4 Network Calculus

5 Optimization

6 Evaluation

7 DRR vs WRR

8 Conclusion


Network Calculus

Network calculus

Computes upper bounds on:

End-to-end delay.Jitter.

Pessimism: models network based on traffic envelops.

Overestimated flow traffic.Underestimated network service.


Network Calculus: Traffic Envelops

System

Network

(DRR)

End

System

End


Network Calculus: Traffic Envelops : Arrival Traffic

System

Network

(DRR)

End

System

End

Actual Arrival Flow

t(µsec)

bits

Smax

f1 f2

BAG


Network Calculus: Traffic Envelops : Arrival Curve

System

Network

(DRR)

End

System

End

t(µsec)

bits

Smax

f1 f2

BAG

αhC1

Smax

f1 f2

BAG


Network Calculus: Traffic Envelops : Network Service

System

Network

(DRR)

End

System

End

Actual Network Service

t(µsec)

bits

Smax

f1 f2

BAG

αhC1

Smax

f1 f2

BAG


Network Calculus: Traffic Envelops : Service Curve

System

Network

(DRR)

End

System

End

t(µsec)

bits

Smax

f1 f2

BAG

αhC1

Smax

f1 f2

BAG

ρhC1

βhC1

ΘhC1


Network Calculus: Traffic Envelops : Delay

System

Network

(DRR)

End

System

End

t(µsec)

bits

Smax

f1 f2

BAG

αhC1

Smax

f1 f2

BAG

ρhC1

βhC1

ΘhC1

delay


Network Calculus: Traffic Envelops : Optimization

System

Network

(DRR)

End

System

End

t(µsec)

bits

Smax

f1 f2

BAG

αhC1

Smax

f1 f2

BAG

ρhC1

βhC1

ΘhC1

delay

β‘hC1

ρhC1

βhC1


Network Calculus: Traffic Envelops : Optimization

System

Network

(DRR)

End

System

End

t(µsec)

bits

Smax

f1 f2

BAG

αhC1

Smax

f1 f2

BAG

ρhC1

βhC1

ΘhC1

delay

β‘hC1

ρhC1

βhC1

α‘hC1

delay

αhC1

delay‘


Table of Contents

1 Context


3 DRR Algorithm

4 Network Calculus

5 Optimization

6 Evaluation

7 DRR vs WRR

8 Conclusion


Pessimism in computed network service

DRR

v1v2

v7

v3v4

v8v9v10

C1 v5v6

C2 v11v12

v13v14v15v16C3 v17v18

v7 v13v14v8

v9 v15v16v10

v11 v17v18v12

v1

v2, v3

v4, v5, v6

βhC1

ΘhC1 Dh

C1

αhC1

t(µsec)

bits


Pessimism in computed network service

DRR

v1v2

v7

v3v4

v8v9v10

C1 v5v6

C2 v11v12

v13v14v15v16C3 v17v18

v7 v13v14v8

v9 v15v16v10

v11 v17v18v12

v1

v2, v3

v4, v5, v6

DRR

v1v2

v7

v3v4

v8v9v10

C1 v5v6

C2 v11

v13v14v15C3

v7 v13v14v8

v9 v15v10

v11

v1

v2, v3

v4, v5, v6

βhC1

ΘhC1 Dh

C1

βhC1

ΘhC1 D‘hC1

DhC1

Pessimism

αhC1

αhC1

t(µsec)

bits

t(µsec)

bits


Upper bound on interfering traffic

DRR

v1v2

v7

v3v4

v8v9v10

C1 v5v6

C2 v11v12

v13v14v15v16C3 v17v18

v7 v13v14v8

v9 v15v16v10

v11 v17v18v12

v1

v2, v3

v4, v5, v6

DRR

v1v2

v7

v3v4

v8v9v10

C1 v5v6

C2 v11

v13v14v15C3

v7 v13v14v8

v9 v15v10

v11

v1

v2, v3

v4, v5, v6

βhC1

ΘhC1 Dh

C1

βhC1

ΘhC1 D‘hC1

DhC1

Pessimism

αhC1

αhC1

t(µsec)

bits

t(µsec)

bits

(αhC2

(DC1))

(αhC3

(DC1))

− = ExtraloadC2Actual traffic from C2 before DhC1

C2 Flow traffic in βhC1




More accurate delay computation

DRR

v1v2

v7

v3v4

v8v9v10

C1 v5v6

C2 v11v12

v13v14v15v16C3 v17v18

v7 v13v14v8

v9 v15v16v10

v11 v17v18v12

v1

v2, v3

v4, v5, v6

DRR

v1v2

v7

v3v4

v8v9v10

C1 v5v6

C2 v11

v13v14v15C3

v7 v13v14v8

v9 v15v10

v11

v1

v2, v3

v4, v5, v6

βhC1

ΘhC1 Dh

C1

βhC1

ΘhC1 D‘hC1

DhC1

Pessimism

αhC1

αhC1

t(µsec)

bits

t(µsec)

bits

(αhC2

(DC1))

(αhC3

(DC1))





Dopt,hC1

= DhC1− tr{ExtraloadC2 + ExtraloadC3}

Dopt,hC1


Table of Contents

1 Context


3 DRR Algorithm

4 Network Calculus

5 Optimization

6 Evaluation

7 DRR vs WRR

8 Conclusion


Evaluation

Airbus A350 configuration

984 flows, 96 end systems, 8 switches, 6412 paths

0

20

40

60

80

100

0 1000 2000 3000 4000 5000 6000 7000

flow paths

DRR_SER normalized to 100DRR_SER_OPT compared to DRR_SER

Average gain:48%.

Max gain 75%.

Class Flows Smax (byte) Qx CategoryC1 718 475 2×Lmax CriticalC2 194 971 Lmax MultimediaC3 72 1535 Lmax Best-effort


Table of Contents

1 Context


3 DRR Algorithm

4 Network Calculus

5 Optimization

6 Evaluation

7 DRR vs WRR

8 Conclusion


Performance Analysis

classNo. of DRR WRR weight frame sizeflows Quantum (bytes) (no. of packets) range

C1 718 4 x lmax 4 415-475C2 194 2 x lmax 2 911-971C3 72 1 x lmax 1 1475-1535

classDRR vs WRR

avg difference (%) max difference (%)

C1 29.16 52.7C2 29.6 52.3C3 -35.4 -68.8


Table of Contents

1 Context


3 DRR Algorithm

4 Network Calculus

5 Optimization

6 Evaluation

7 DRR vs WRR

8 Conclusion


Conclusion

NC on AFDX network with mixed criticality

QoS: DRR scheduling.

Evaluation of improved NC approach.

Performance comparison of DRR and WRR schedulers.

What’s next?

Exact worst case delay using model checking approach.Evaluation of pessimism of NC for avionic network withDRR and WRR scheduler.Weight/Quantum allocation in Round Robin scheduler(WRR/DRR)


Questions

Thank you for your attention!

Aakash SONI - [email protected]


optimizing network calculus for switched ethernet network...

Documents