Title: Statistical Modeling for PerHop QoS
1Statistical Modeling for Per-Hop QoS
- Mohamed El-Gendy
- (mgendy_at_eecs.umich.edu)
- In collaboration with
- Abhijit Bose, Haining Wang, and Prof. Kang G.
Shin - Real-Time Computing Laboratory
- EECS Department
- The University of Michigan_at_Ann Arbor
- June 4th, 2003
2Outline
- Intro of DiffServ, PHB, and Per-Hop QoS
- Motivations
- Related Work
- Approach to Statistical Characterization
- Experimental Framework
- Results and Analysis
- A Control Example
- Conclusions and Future Work
3DiffServ and PHB
- Scalable network-level QoS based on marked
traffic aggregates - Traffic is conditioned and marked with DSCP at
edge - Per-Hop Behaviors (PHBs) are applied to traffic
aggregates at core - QoS is achieved through different PHBs
- Expedited Forwarding (EF) for delay assurance
- Assured Forwarding (AF) for bandwidth assurance
4Per-Hop QoS
- Throughput (BW), delay (D), jitter (J), and loss
(L) experienced by traffic crossing a PHB
5Motivations
- Why modeling Per-Hop QoS?
- PHB is the key building block of DiffServ
- Wide variety of PHB realizations
- PHB control and configuration
- Necessary for end-to-end QoS calculation
- Benefits of PHB modeling
- Facilitates the control and optimization of PHB
performance - Enables contribution of per-hop admission control
to e2e admission control decisions
6Related Work
- Study of TCP ACK marking in DiffServ IWQoS01
- Used full factorial design and ANOVA
- Compared many marking schemes for TCP acks
- Suggested an optimal strategy for marking the
acks for both assured and premium flows - Used ns simulation for analysis
- AF performance using ANOVA IETF draft
- Compared different bandwidth and buffer
management schemes for their effect on AF
performance
7Related Work
- Performance of TCP Vegas Infocom00
- Used ANOVA to test the effect of ten congestion
and flow control algorithms - Clustered the ten factors into three groups
according to the the three phases of the TCP
Vegas operation
8Approach to Statistical Characterization
- Identify the factors in I and C that affect
output per-hop QoS most - Construct statistical models of the per-hop QoS
in terms of these important factors
9Input Traffic Factors - I
- Dual Leaky Bucket (DLB) representation for I
- Average rate, peak rate, burst size, packet size,
number of flows per aggregate, and traffic type - Ia assured traffic, Ib background traffic
- Used the ratio between assured to best-effort
traffic, instead of absolute value - Number of input interfaces to PHB node
10Alternative PHB Realizations
EF-EDGE
EF-CORE
EF-CBQ
- Different PHB realizations have different
functional relationships between inputs and
outputs
11Configuration Parameters- C
- Configuration parameters depend on PHB
realization - EF-EDGE token rate, bucket size, and MTU
- EF-CORE queue length
- EF-CBQ service rate, burst size, and avg packet
size - AF min. threshold, max. threshold, and drop
probability
12Statistical Analysis
- Analysis of Variance (ANOVA)
- Models
- Output response as a linear combination of the
main effects and their interactions - Allocation of variation
- Calculate the percentage of variation in the
output response due to factors at each level,
their interactions, and the errors in the
experiments - ANOVA
- Statistically compare the significance of each
factor as well as the experimental error
13ANOVA
- For any three factors (k 3), A, B, and C with
levels a, b, and c, and with r repetitions, the
response variable y can be written as
14ANOVA
- Squaring both sides we get
SSE sum of squared errors
15ANOVA
16ANOVA Model Assumptions
- Assumptions
- Effects of input factors and errors are additive
- Errors are identical, independent, and normally
distributed random variables - Errors have a constant standard deviation
- Visual tests
- No trend in the scatter plot of residuals vs.
predicted response - Linear normal quantile-quantile (Q-Q) plot of
residuals
No assumptions about the nature of the
statistical relationship between input factors
and response variables
17Statistical Analysis - Regression
- Polynomial regression
- A variant of multiple linear regression
- Any complex function can be expanded into
piecewise polynomials with enough number of terms - Transformations to deal with nonlinear dependency
- Coefficient of determination (R2) as a measure of
the regression goodness
18Regression
Linear model for one dependent variable y and k
independent variables x
Polynomial model for two independent variables
x1, x2
Transformation to fit into linear model
19Experimental Framework
- Framework components
- Traffic Generation Agent
- Generates both TCP and UDP traffic
- Policed with a built-in leaky bucket for profiled
traffic - BW, D, J, L are measured within the agent itself
- Controller and Remote Agents
- Control the flow of the experiments according to
a distributed scenario file - Executes and keep track of the experiments steps
and other components
20Experimental Framework, contd
- Network and Router Configuration Agents
- Configure traffic control blocks on router
according to experiment scenarios - Receive scenario commands from the Controller
agent - Current implementation works on Linux traffic
control
- Analysis Module
- Performs ANOVA, model validation tests, and
polynomial regression on output data
21Experimental Framework, contd
- Network Setup
- Using ring topology for one-way delay measurements
22Full Factorial Design of Experiments
- If we have k factors, with ni levels for the
i-th factor, and repeat r times - Total number of experiments
LARGE!!
- Use factor clustering and automated experimental
framework
23Scenarios of Experiments
- EF PHB
- Factor sets Ia, Ib, C
- PHB configurations EF-EDGE, EF-CORE, EF-CBQ
- Operating mode over-provisioned (OP),
under-provisioned (UP), fully-provisioned (FP)
24Scenarios of Experiments, contd
- AF PHB
- Use AF11 as assured traffic
- Use AF12 and AF13 as background traffic
- Change max. threshold, min. threshold, and drop
probability for AF11 only
25EF PHB OP, EF-EDGE w/o BG traffic
- BW surface response significant factors are
assured rate ( ar ) and number of assured flows (
an ), R2 96
26EF PHB OP, EF-EDGE w/o BG traffic
Significant factors are assured rate ( ar ),
number of assured flows ( an ), and assured
packet size (apkt)
27EF PHB UP, EF-EDGE w/o BG traffic
- ANOVA results for L
- Significant factors are assured rate ( ar ),
number of assured flows ( an ), and the token
bucket rate ( efr )
28EF PHB OP, EF-EDGE w/ BG traffic
- ANOVA results for BW, D, and J
- Significant factors are BG packet size ( bpkt ),
number of BG flows ( bn ), and ratio of assured
to BG traffic ( Rab )
29EF PHB EF-CORE w/o BG traffic
- J surface response
- Significant factors are assured packet size (
apkt ) and number of assured flows ( an ), R2
64
30EF PHB EF-CORE w/o BG traffic
31EF PHB OP, EF-CBQ w/o BG traffic
- ANOVA results for BW, D, and J
- Significant factors are assured rate ( ar ),
assured packet size ( apkt ) and number of
assured flows ( an )
32EF PHB OP, EF-CBQ w/ BG traffic
- ANOVA results for L
- Significant factors are BG packet size ( bpkt ),
number of BG flows ( bn ), and ratio of assured
to BG traffic ( Rab )
33EF PHB OP, EF-CBQ w/ BG traffic
34AF PHB
- ANOVA results for BW, D, J, and L
- Significant factors are assured rate ( ar ),
assured peak rate ( ap ), assured packet size (
apkt ), max. threshold ( maxth ) , and min.
threshold ( minth )
35Discussion
- BW shows a square root relationship with factors
in Ia in EF-CBQ only, and direct relation in the
other EF realizations
- D shows a direct relation with Ia in EF-EDGE, and
EF-CORE, and inverse relation in EF-CBQ
- D shows a logarithmic (multiplicative) relation
with Ib
- J shows inverse relation with Ia and a direct
relation with Ib
- J depends on the number of flows in the aggregate
as well as the difference in packet size with
other flows/aggregates
36Errors
- Experimental errors due to experimental methods
captured in ANOVA - Model errors due to factor truncation
- Statistical and fitting errors due to
regression captured in coefficient of
determination (R2 )
37A PHB Control Example
- For OP, EF-CBQ w/ BG traffic
- For bpkt 600 B, bn 1, Rab 2 ? D 0.4136
msec - For bpkt 1470 B, bn 3, D 0.4136 msec ? Rab
?? - Use the delay model to find Rab 0.494 with
accuracy of (1-R2) 11
38Conclusions
- Simple statistical models are derived for per-hop
QoS using ANOVA and polynomial regression - Statistical full factorial design of experiments
is an effective tool for characterizing QoS
systems - Using automated experimental framework is shown
to be effective in such studies - Different PHB realizations show differences in
dependency of per-hop QoS on input factors
39Extensions and Future Work
- More rigorous control analysis and study of
suitable control algorithms
- Validate the models derived with analytical
methods such as network calculus
- Use real-time measurements to update models and
control criterion while operation
- The framework presented is general to be applied
for studying edge-to-edge (Per-Domain Behavior or
PDB) in DiffServ
40Multi-Hop Case
41Multi-Hop Case
42Questions ?