Layering, dynamics, optimization & control in software defined - - PowerPoint PPT Presentation

Layering, dynamics, optimization & control in software defined networks

Nikolai Matni Computing and Mathematical Sciences

joint work with John Doyle (Caltech) and Kevin Tang (Cornell)

Nikolai Matni Computing and Mathematical Sciences

joint work with John Doyle (Caltech) and Kevin Tang (Cornell)

Distributed optimal control & software defined networks

3

Control theory

plant set-point disturbances “open loop” Using feedback to mitigate the effects

• f dynamic uncertainty on a system

4

Control theory

plant set-point Sensors Controller Actuators

measurements control action

feedback to close the loop disturbances

5

A familiar example: TCP

Feedback: AIMD based on drops, RTT, queue length, etc. Guarantees: converge to NUM optimizing transmission rates

Only steady state guarantees

converge

6

Stabilizing controllers

Controller #1 Controller #2

7

Optimal control theory

amplification

error control action disturbance

minimize worst-case

Bounded energy Bounded magnitude

8

Optimal control theory

white noise

amplification

error control action disturbance

minimize average

Centralized control

Sensors Plant Actuators Controller control inputs measurements

One system One controller Global access to measurements Global control

• f inputs

Can lead to poor performance for large-scale systems

9

Distributed control

S P A C

10

Distributed control

11

S P A C S P A C S P A C S P A C S P A C S P A C S P A C S P A C S P A C S P A C S P A C S P A C S P A C S P A C S P A C S P A C S P A C S P A C

convex!

if control packets get priority

Distributed optimal control in WANs

12

TE solved using nominal demands

WAN distributed optimal control

13

queue length egress buffer control

High Frequency Traffic Control

rate deviation

L1 S1

“ ” fl fl fl fi fl fl λ fl

∗

fl ∆ –

real traffic fluctuates around nom rates

fi τ – §

∗ σ

σ

∗

fl fl → → → –

S2 S1 S3 20ms Src2 Src1 Dst1,2

µ, σ σ µ ↵ fl fl ↵ ↵ ↵ ↵ ↵ λ ↵ ↵ λ ↵ fi ’ ↵ λ λ ↵ fi ↵ ↵ λ ↵

® ¥

↵ ↵ ↵ λ ↵ ↵ ↵ ↵ ↵ fl ↵

FIFO Smoothing

Packet loss (%) Average queue length (Mbits)

• F

Generalizes FIFO/Smoothing

egress buffer control queue length rate deviation

14

Controller architectures

15

queue length egress buffer control

High Frequency Traffic Control

rate deviation

Globally Optimal Delay Free (GOD-F)

controller

0 delay comms

Controller architectures

16

queue length egress buffer control

High Frequency Traffic Control

rate deviation

Centralized

controller

Controller architectures

17

queue length egress buffer control

High Frequency Traffic Control

rate deviation

Distributed

control control control control

new theory

ctrl packets get priority = convex!

Controller architectures

18

queue length egress buffer control

High Frequency Traffic Control

rate deviation

Decentralized (myopic)

control control control control

non-convex use best guess

FIFO (standard) centralized decentralized distributed

Max link utilization most delay least delay centralized decentralized

WAN reflex layer

Demand fluctuation std. dev.

19

• N. Wu, A. Tang, J.C. Doyle, …, N. Matni, in preparation
• N. Matni, A. Tang, J.C. Doyle, ACM SIGCOMM Symposium on SDN Research, 2015

σ= 5

σ= 10

theory & experiments are consistent!

A theory of network architecture

20

• deciding where to put functionality?
• layering and inter/intra-layer protocol design?
• choosing centralized, distributed
• r decentralized implementations?

Can we understand & automate:

Must model dynamics & delay

Nikolai Matni

Select references:

• N. Matni & J. C. Doyle, A theory of dynamics, control and optimization in layered

architectures, IEEE American Control Conference, 2016

• N. Matni, A. Tang & J. C. Doyle, A case study in network architecture tradeoffs,

ACM Symposium on SDN Research (SOSR), 2015 High-frequency traffic control preprint available upon request.