http://www.ece.rice.edu/networks Aleksandar Kuzmanovic and Edward W. Knightly Rice Networks Group Measuring Service in Multi- Class Networks
Dec 17, 2015
http://www.ece.rice.edu/networks
Aleksandar Kuzmanovic and Edward W. KnightlyRice Networks Group
Measuring Service in Multi-Class Networks
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Background
QoS services– SLA guaranteed rate
Ex. Class X serviced at minimum rate R
– Relative performance Ex. Class X has strict
priority over class Y
– Statistical service Ex. P(class X pkt.
Delay>100ms)<.001
QoS mechanisms– Priority queues
Rate-based, delay-based...
– Policing Rate limiting...
– Over-engineering Just add more
bandwidth...
Need: Tools for network clients to assess the networks QoS capabilities
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Inverse QoS Problem
Is a class rate limited? What is the inter-class relationship?
– Fair/weighted fair/strict priority Is resource borrowing fully allowed or not? Is the service’s upper bound identical to its lower
bound? What are the service’s parameters?
A ssess m echa n ism s a nd pa ra m eters o f a n u nk now n Q oS system
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Applications - Network Example
Providers reluctant to divulge precise QoS policy (if any...)
SLA validation for VPNs– Is the SLA fulfilled?
Capacity planning– What is the relationship
among classes?
Edge-based admission control [CK00] and implementation [SSYK01]
A
B
V PN cla s s 1V PN cla s s 2B a ck g ro u n d
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Performance Monitoring and Resource Management
Single WEB server– CPU resource sharing– Listen queue differentiation– Admission control
Distributed WEB server– Load balancing
Internet Data Center– Machine migration
F ro nt EndS erver
B ac k-endS erver
M eas urem entM o d ule
B ac k-endS erver
B ac k-endS erver
Goal: Estimate a class’ net “guaranteed rate”
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
“Off-Line” Solution is Simple
Consider a router with unknown QoS mechanisms
U nkno w n Q o S M e c hani s m
I n p u t O u tp u t
C las s 1
C las s 2
P a c k e t A rriv a ls
O u tp u t R a te
C la s s 1ra te lim ite d
C la s s 2no t ra te lim ite d
W e ighte d F a irne s s
F ull C ap ac ity
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
“On-Line” Case: Operational Network
Undesirable to disrupt on-going services– High rate probes to detect inter-class
relationships would degrade performance Impossible to force other classes to be idle
– … to detect policers
U nkno w n Q o S M e c hani s m
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
System Model and Problem Formulation
Two stage server– Non-work conserving elements– Multi-class scheduler
Observations– Arrival and
departure times– Class ID– Packet size
R a te L im ite rs U n kn o w n M u lti-C la s s S e rve r
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Determine...
Infer the service discipline– Most likely hypothesis among WFQ, EDF and SP
Detect the existence of non-work conserving elements– Rate limiters (ex. leaky bucket policers)
Estimate the system parameters– WFQ guaranteed rates, EDF deadlines, rate
limiter values
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Remaining Outline
Inter-class Resource Sharing Theory
Empirical Arrival and Service Models
MLE of Parameters
EDF/WFQ/SP Hypothesis Testing
Simulation Results and Conclusions
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Theoretical Tool: Statistical Service Envelopes [QK99]
General statistical char. for a (virtual) minimally backlogged flow
Flows receive additional service beyond min rate– Function of other flow
demand– Function of scheduler
General characterization of inter-class resource sharing
Framework for admission control for EDF/WFQ/SP
in terval
serv
ice
guaranteedrate
99% s erv ic e
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Inter-class theory
Key technique:– Passively monitor arrivals and services at edges– Devise hypothesis tests to jointly:
Detect most likely hypothesis Estimate unknown parameters
Strategy
),),(()( kni HtBftS
k
n
i
H
tB
tS
)(
)( S e rv ice
A rriv a ls
H y po th e s is
Un k n o wn s
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Empirical Arrival Model
Envelopes characterize arrivals as a function of interval length– Statistical traffic envelope [QK99]
Empirical envelope - measure first two moments of arrivals over multiple time scales
time
t + It
E*( I ) = 3
Goal: assuming Gaussian distribution for B
),|(),,( ki
kni HSpHBfS
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Empirical Service Model
A real-world paradigm for statistical service envelope
Observe: Service can be measured only when packets are backlogged
A rriv a ls
D e p a rtu re s
Ser
vice
I n te rv a l
A rriv a ls
D e p a rtu re sS
ervi
ce
I n t e rv a l
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Empirical Service Distributions
For each class and time scale– Expected service distributions– Service measures (data)
Empirical service distributions
),|( kHsp
WFQ (400 ms) SP (400 ms)
0 100 200 300 400 500 600 700 800 9000
5
10
15
20
25
30
35
Service rate (Kbps)
Em
piric
al r
ate
pro
abili
ty
0 100 200 300 400 500 600 700 800 9000
5
10
15
20
25
30
35
Service rate (Kbps)
Em
piric
al r
ate
pro
abili
ty
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Parameter Estimation andScheduler Inference
GLRT for each time scale
Under MLE parameters for
each scheduler Choose most likely scheduler Apply majority rule over all
time scales
),,|,(max
),,|,(max),(
~
2121,
2121,
21
21
21
ddEDFssp
WFQsspss
dd
><1
i
i
i
d
s
Se r vi c e fo r c l as s i (data)
H ypo the s i s 1
H ypo the s i s 2
U nkno w n par am e te r s
W FQ
E D F
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
EDF/WFQ Testing
Correctness ratio
True WFQ 94%
True EDF 100%
Importance of time scales
Short time scales– Fluid vs. packet model
Long time scales– Ratio of delay shift and
time scale decreases as time scale increases (d1=25ms)
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Measurable Regions
What if there is no traffic in particular class?
What traffic load “allows” inferences?
Region where we are able to estimate true value within 5%
Typical utilization should be > 62% for 1.5 Mbps link
Otherwise, active probing required
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Conclusions
Framework for clients of multi-class services to assess a system’s core QoS mechanisms
– Scheduler type
– Estimate parameters (both w-c and n-w-c)
General multiple time-scale traffic and service model to characterize a broad set of behaviors within a unified framework
http://www.ece.rice.edu/networks
Aleksandar Kuzmanovic and Edward W. KnightlyRice Networks Group
Measuring Service in Multi-Class Networks
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Ongoing Work
Unknown cross-traffic– Cannot monitor all
systems inputs/outputs– Treat cross-traffic statistics
as another unknown Web servers
– Evaluation of the framework in a single web server through trace driven simulations
– Capacity is statistically characterized
A
B
V PN cla s s 1V PN cla s s 2B a ck g ro u n d
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
WFQ Parameter Estimation
Class 1: 65-68 flows Class 2: 25-28 flows Large windows improve
confidence level– T=2sec: 95% in 11% of
true value– T=10sec: 95% in 1.4% of
true value
Flow level dynamics & non-
stationarities must be
considered
1 2 3 4 5 6 7 8 9 10 110.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
Measurement interval (sec)
WF
Q r
elat
ive
wei
ght
estim
ate
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Rate Limited Class State Detection
Can include parameter r in service envelope equations for each class
Importance of time scales
Example– Class based fair queuing– C=1.5Mbps, r=1Mbps
Probability decreases with time scale higher errors when measuring multi-level leaky-buckets
Kuzmanovic & Knightly | Rice Networks Group | INFOCOM 2001
Generalized Likelihood Ratio Test
Detection with unknowns
Note: we do not find a single value of that maximizes likelihood ratio
Under mild conditions (as ), GLRT is Uniformly Most Powerful (maximizes the probability of detection)
),|(max
),|(max)(
~
00
11
0
1
Hxp
Hxpx ><
1H
0H1
ji
iH
x
,
D ata s e t
H ypo the s i s
U nkno w n par am e te r s
N