Feedback Control Theory

a Computer System’s Perspective

Chenyang Lu
Department of Computer Science
University of Virginia
chenyang@cs.virginia.edu
http://www.cs.virginia.edu/~cl7v
Outline
 Introduction
 What is feedback control?
 Why do computer systems need feedback control?
 Control design methodology
 System modeling
 Performance specs/metrics
 Controller design
 Summary
Control

 Applying input to cause system variables to

conform to desired values called the reference.
 Cruise-control car: f_engine(t)=?  speed=60 mph
 E-commerce server: Resource allocation?  T_response=5 sec
 Embedded networks: Flow rate?  Delay = 1 sec
 Computer systems: QoS guarantees
Open-loop control

 Compute control input without continuous variable

measurement
 Simple
 Need to know EVERYTHING ACCURATELY to work right
 Cruise-control car: friction(t), ramp_angle(t)
 E-commerce server: Workload (request arrival rate? resource
consumption?); system (service time? failures?)
 Open-loop control fails when
 We don’t know everything
 We make errors in estimation/modeling
 Things change
Feedback (close-loop) Control

Controlled System
Controller
control manipulated
control Actuator
input variable
function

error
sample controlled
+ Monitor
- variable

reference
Feedback (close-loop) Control
 Measure variables and use it to compute control input
 More complicated (so we need control theory)
 Continuously measure & correct
 Cruise-control car: measure speed & change engine force
 Ecommerce server: measure response time & admission control
 Embedded network: measure collision & change backoff window
 Feedback control theory makes it possible to control
well even if
 We don’t know everything
 We make errors in estimation/modeling
 Things change
Why feedback control?
Open, unpredictable environments

 Deeply embedded networks: interaction with

physical environments
 Number of working nodes
 Number of interesting events
 Number of hops
 Connectivity
 Available bandwidth
 Congested area
 Internet: E-business, on-line stock broker
 Unpredictable off-the-shelf hardware
Why feedback control?
We want QoS guarantees

 Deeply embedded networks

 Update intruder position every 30 sec
 Report fire <= 1 min
 E-business server
 Purchase completion time <= 5 sec
 Throughput >= 1000 transaction/sec
 The problem: provide QoS guarantees in
open, unpredictable environments
Advantage of feedback
control theory

 Adaptive resource management heuristics

 Laborious design/tuning/testing iterations
 Not enough confidence in face of untested workload
 Queuing theory
 Doesn’t handle feedbacks
 Not good at characterizing transient behavior in
overload
 Feedback control theory
 Systematic theoretical approach for analysis and design
 Predict system response and stability to input
Outline
 Introduction
 What is feedback control?
 Why do today’s computer systems need feedback control?
 Control design methodology
 System modeling
 Performance specs/metrics
 Controller design
 Summary
Control design methodology

Controller
Modeling Design
analytical Dynamic model Root-Locus PI
Control algorithm
system IDs Control

Satisfy

Requirement
Performance Specifications
Analysis
System Models

 Linear vs. non-linear (differential eqns)

 Deterministic vs. Stochastic
 Time-invariant vs. Time-varying
 Are coefficients functions of time?
 Continuous-time vs. Discrete-time
 System ID vs. First Principle
Dynamic Model
 Computer systems are dynamic
 Current output depends on “history”
 Characterize relationships among system variables
• Differential equations (time domain)
  
a2 y (t )  a1 y (t )  a0 y (t )  b1 u (t )  b0u (t )
• Transfer functions (frequency domain)
Y(s) = G(s)U(s)
b1s  b0 c1 c2
G ( s)   
a2 s 2  a1s  a0 s  p1 s  p2
• Block diagram (pictorial)

R(s) C(s) G(s) Y(s)

-
Example
Utilization control in a video server
 Periodic task Ti corresponding to each video stream i
 c[i]: processing time, p[i]: period
 Stream i’s requested CPU utilization: u[i]=c[i]/p[i]
 Total CPU utilization: U(t)={k}u[k], {k} is the set of
active streams
 Completion rate: Rc(t)= ({kc}u[m])/t, where {m} is the
set of terminated video streams during [t, t+t]
 Unknown
 Admission rate: Ra(t)= ({ka}u[j])/t, where {j} is the set
of admitted streams during [t, t+t]
 Problem: design an admission controller to guarantee
U(t)=Us regardless of Rc(t)
Model
Differential equation

• Error: E(t)=Us-U(t)
t

• Model (differential equation): U (t )   ( R ( )  R ( ))d

a c
 0

• Controller C? E(t)  Ra(t)

Ra(t)
- Us
C?
U(t)
CPU

Rc(t)
A Diverge to Math
System representations

 Three ways of system modeling

• Time domain: convolution; differential equations.
t

u(t) g(t) y(t) y (t )  g (t ) * u (t )   g (t   )u ( )d

0

• s (frequency) domain: multiplication

U(s) G(s) Y(s) Y ( s )  G ( s )U ( s )

• Block diagram: pictorial

s-domain is a simple & powerful “language” for control analysis

A Diverge to Math
Laplace transform

 Laplace transform of a signal f(t)



F ( s )  L[ f (t )]   f (t )e  st
dt
0

where s=+i is a complex variable.

• Laplace transform is a translation from time-domain to
s-domain
• Differential equation  Polynomial function
  
a2 y (t )  a1 y (t )  a0 y (t )  b1 u (t )  b0u (t )
b1s  b0
 Y (s)  U ( s)
a2 s  a1s  a0
2
A Diverge to Math
Laplace transform

 Basic translations
 Impulse function f(t)=(t)  F(s)=1
 Step signal f(t)=a•1(t)  F(s)=1/s
 Ramp signal f(t)=a•t  F(s)=a/s2
 Exp signal f(t)=eat  F(s)=1/(s-a)
 Sinusoid signal f(t)=sin(at)  F(s)=a/(s2+a2)
 Composition rules
 Linearity L[af(t)+bg(t)] = aL[f(t)]+bL[g(t)]
 Differentiation L[df(t)/dt] = sF(s) – f(0-)
 Integration L[tf()d] = F(s)/s
A Diverge to Math
Transfer function

 Modeling a linear time-invariant (LTI)

system
 G(s) = Y(s)/U(s)  Y(s) = G(s)U(s)

U(s) G(s) Y(s)

E.g., a second order system with poles p1 and p2

b1s  b0 c1 c2
G(s)   
a2 s 2  a1s  a0 s  p1 s  p2
A Diverge to Math
Poles and Zeros

 The response of a linear time-invariant (LTI) system

bm s m  bm1s m1  ...  b0
F (s) 
an s n  an1s n1  ...  a0
im1 ( s  zi ) C1 C2 Cn
K n    ... 
i 1 ( s  pi ) s  p1 s  p2 s  pn
n
 f (t )   Ci e pit
i 1

{pi} are poles of the function and decide the system behavior
A Diverge to Math
Time response vs. pole location
Stable Unstable

•f’(t) = ept, p = a+bj

A Diverge to Math
Block diagram
 A pictorial tool to represent a system based on transfer
functions and signal flows
 Represent a feedback control system

R(s) C(s) Go(s) Y(s)

-

C ( s )Go ( s )
Gc 
R(s) Gc(s) Y(s) 1  C ( s )Go ( s )
Y ( s )  Gc ( s ) R ( s )
Back to
Our utilization control example

• Error: E(t)=Us-U(t)
t

• Model (differential equation): U (t )   ( R ( )  R ( ))d

a c
 0

• Controller C? E(t)  Ra(t)

Ra(t)
- Us
C?
U(t)
CPU

Rc(t)
Model
Transfer func. & block diag.

• CPU is modeled as an integrator

t
Ra ( s )  Ra ( s ) 1
U (t )   ( R ( )  R ( ))d  U (s) 
a c
s
 Go ( s) 
s
 0

 Inputs: reference Us(s) = Us/s; completion rate Rc(s)

 Close-loop system transfer functions
 Us(s) as input: G1(s) = C(s)Go(s)/(1+C(s)Go(s))
 Rc(s) as input: G2(s) = Go(s)/(1+C(s)Go(s))
 Output: U(s)=G1(s)Us/s+G2(s)Rc(s)
Rc(s)

Ra(s) Go
Us/s C(s) U(s)
Control design methodology

Controller
Modeling Design
analytical Dynamic model Root-Locus PI
Control algorithm
system IDs Control

Satisfy

Requirement
Performance Specifications
Analysis
Design Goals
Performance Specifications

 Stability
 Transient response
 Steady-state error
 Robustness
 Disturbance rejection
 Sensitivity
Performance Specs
Stability

 BIBO stability: bounded input results in bounded output.

 A LTI system is BIBO stable if all poles of its transfer function
are in the LHP (pi, Re[pi]<0).

im1 ( s  zi ) C1 C2 Cn
Y ( s )  G ( s )U ( s )  K n    ... 
i 1 ( s  pi ) s  p1 s  p2 s  pn
n
 y (t )   Ci e pi t
i 1
t 
Note : Ci e pi t 
  if Re[ pi ]  0
Performance Specs
Stability

Stable Unstable
 Performance specifications
Controlled
variable
Overshoot
Steady state error
%
Reference

Transient State Steady State

Time
Settling time
 Example: Control & Response
in an Email Server (IBM)
Response
(queue length)

Good
Control Bad
(MaxUsers)

Slow Useless
Performance Specs
Steady-state error

 Steady state (tracking) error of a stable system

ess  lim e(t )  lim(r (t )  y (t ))

t  t 

r(t) is the reference input, y(t) is the system output.

• How accurately can a system achieve the desired state?
• Final value theorem: if all poles of sF(s) are in the open left-half
of the s-plane, then

lim f (t )  lim sF ( s )
t  s 0

• Easy to evaluate system long term behavior without solving it

ess  lim e(t )  lim sE ( s)

t  s 0
Performance Specs
Steady-state error

Steady state error of a CPU-utilization control system

U(t)

Us
ess=-20%
Performance Specs
Robustness

 Disturbance rejection: steady-state error caused by

external disturbances
 Can a system track the reference input despite of external
disturbances?
 Denial-of-service attacks
 Sensitivity: relative change in steady-state output
divided by the relative change of a system parameter
 Can a system track the reference input despite of variations in
the system?
 Increased task execution times
 Device failures
Performance Specs
Goal of Feedback Control

 Guarantee stability
 Improve transient response
 Short settling time
 Small overshoot
 Small steady state error
 Improve robustness wrt uncertainties
 Disturbance rejection
 Low sensitivity
Control design methodology

Controller
Modeling Design
analytical Dynamic model Root-Locus
Control algorithm
system IDs PID Control

Satisfy

Requirement
Performance Specifications
Analysis
Controller Design
PID control

 Proportional-Integral-Derivative (PID) Control

• Proportional Control x(t )  Ke(t )  C (s)  K
t
KK i
• Integral control x(t )  KK i  e( )d  C (s) 
0 s

• Derivative control x(t )  KK d e (t )  C ( s )  KK d s

• Classical controllers with well-studied properties and tuning rules

E(s) X(s)
R(s) C(s) Go(s) Y(s)
-
Controller Design
CPU Utilization Control

• CPU is modeled as an integrator

t
Ra ( s )  Ra ( s ) 1
U (t )   ( R ( )  R ( ))d  U (s) 
a c
s
 Go ( s) 
s
 0

 Inputs: set-point Us(s) = Us/s ; task completion Rc(s)

 Close-loop system transfer functions
 Us(s) as input: G1(s) = C(s)Go(s)/(1+C(s)Go(s))
 Rc(s) as input: G2(s) = Go(s)/(1+C(s)Go(s))
 C(s)=? to achieve zero steady-state error: U(t)  Us
Rc(s)

Ra(s) Go
Us/s C(s) U(s)
Proportional Control
Stability

 Proportional Controller
 ra(t)=Ke(t); C(s) = K
 Transfer functions
 Us/s as input: G1(s) = K/(s+K)
 Rc(s) as input: G2(s) = 1/(s+K)
 Stability
 Pole p0 = -K<0  System is BIBO stable iff K>0
 Note: System may shoot to 100% if K<0!
Rc(s)

Ra(s) Go
Us/s C(s) U(s)
Proportional Control
Steady-state error
 Assume completion rate Rc(t) keeps constant for a time
period longer than the settling time: Rc(s)=Rc/s
 System response is
U s G1 ( s ) Rc G2 ( s ) KU s  Rc
U ( s)   
s s s(s  K )
• Compute steady-state err using final value theorem,
KU s  Rc Rc Rc
limU (t )  lim sU ( s )  lim  Us   ess   0
t  s 0 s 0 sK K K
• P-control cannot achieve the desired CPU utilization Us; instead
it will end up lower by Rc/K Oops!
• The larger the proportional gain K is, the closer will CPU
utilization approach to Us
CPU Utilization
Proportional Control

U(t)

Us
ess=-20%
Proportional-Integral Control
Stability
 Proportional Controller
 ra(t)=K(e(t)+Ki•te()d) C(s) = K(1+Ki/s)
 Transfer functions
 Us/s as input: G1(s) = (Ks+KKi)/(s2+Ks+KKi)
 Rc(s) as input: G2(s) = s/(s2+Ks+KKi)
 Stability
 Poles Re[p0]<0, Re[p0]<0
 System is BIBO stable iff K>0 & Ki>0
Rc(s)

Ra(s) Go
Us/s C(s) U(s)
Proportional Control
Steady-state error
 Assume completion rate Rc(t) keeps constant for a time
period longer than the settling time: Rc(s)=Rc/s
 System response is
U s G1 ( s ) Rc G2 ( s ) ( KU s  Rc ) s  KK iU s
U ( s)   
s s s ( s 2  Ks  KK i )
• Compute steady-state err using final value theorem,
( KU s  Rc ) s  KK iU s
limU (t )  lim sU ( s )  lim  Us  ess  0
t  s 0 s 0 s  Ks  KK i
2

• PI control can accurately achieve the desired CPU utilization Us 

• Control analysis gives design guidance
CPU Utilization
Proportional-Integral Control

U(t)
Mp=0
Us
ess=0

t s tr t p
Controller Design
Summary & pointers
 PID control: simple, works well in many systems
 P control: may have non-zero steady-state error
 I control: improves steady-state tracking
 D control: may improve stability & transient response
 Linear continuous time control
 Root-locus design
 Frequency-response design
 State-space design
 G. F. Franklin et. al., Feedback control of dynamic systems
Discrete Control
 More useful for computer systems
 Time is discrete; sampled system
 denoted k instead of t
 Main tool is z-transform

Z[ f (k )]  F ( z )   f (k ) z  k
k 0
 f(k)  F(z) , where z is complex
 Analogous to Laplace transform for s-domain
Discrete Modeling
 Difference equation
 V(m) = a1V(m-1) + a2V(m-2) + b1U(m-1) + b2U(m-2)
 z domain: V(z) = a1z-1V(z) + a2z-2V(z) + b1z-1U(z) + b2z-2U(z)
 Transfer function G(z) = (b1z + b2)/(z2-a1z - a2)
 V(m): output in mth sampling window
 U(m): input in mth sampling window
 Order n: #sampling-periods in history affects current
performance
 SP = 30 sec, and n = 2  Current system performance
depends on previous 60 sec
Root Locus analysis of
Discrete Systems

 Stability boundary: |z|=1 (Unit circle)

 Settling time = distance from Origin
 Speed = location relative to Im axis
 Right half = slower
 Left half = faster
Effect of discrete poles
Im(s)

Higher-frequency
response
Longer settling time

Re(s)
Stable

Unstable |z|=1

Intuition : z  e Ts
Feedback control works in CS
 U.Mass: network flow controllers (TCP/IP – RED)
 IBM: Lotus Notes admission control
 UIUC: Distributed visual tracking
 UVA
 Web Caching QoS
 Apache Web Server QoS differentiation
 Active queue management in networks
 Processor thermal control
 Online data migration in network storage (with HP)
 Real-time embedded networking
 Control middleware
 Feedback control real-time scheduling
Advanced Control Topics
 Robust Control
 Can the system tolerate noise?
 Adaptive Control
 Controller changes over time (adapts)
 MIMO Control
 Multiple inputs and/or outputs
 Stochastic Control
 Controller minimizes variance
 Optimal Control
 Controller minimizes a cost function of error and control energy
 Nonlinear systems
 Neuro-fuzzy control
 Challenging to derive analytic results
Issues for Computer
Science
 Most systems are non-linear
 But linear approximations may do
 eg, fluid approximations
 First-principles modeling is difficult
 Use empirical techniques
 Mapping control objectives to feedback control
loops
 ControlWare paper
 Deeply embedded networking
 Massively decentralized control problem
 Modelling
 Node failures
Acknowledgements
 UVA
 Jack Stankovic
 Tarek F. Abdelzaher
 Sang H. Son
 Gang Tao
 HP Labs
 Guillermo Alvarez
 IBM TJ Watson
 Joseph Hellerstein
 Sujay Parekh