Transcript output link - Homepages | The University of Aberdeen

Optimization of Scheduling Algorithm Parameters
in a DiffServ Environment
Authors:
Davide Adami
Stefano Giordano
Michele Pagano
Raffaello Secchi
Speaker:
Raffaello Secchi
Network Telecommunication Group
University of Pisa - Information Engineering department
January 31 2005
1
Outline
•Introduction to scheduling algorithms
•Deficit Weighted Round Robin
•Weighted Fair Queuing
•Objective of our study
•Performance Comparison between DRR and WFQ scheduler
•Derivation of a configuration strategy of scheduling parameters to
minimize the end-to-end delay of real-time application in DRR networks
•Numerical Analysis
•Simulation results in high speed networks
•Conclusions
2
Scheduling Algorithms
The weight associated
to the i-th queue is
proportional to the
percentage of output
capacity
W1
Server
W2
s
Output link
W3
• In this work we considered two different proportional schemes
•Deficit Round Robin (frame-based scheduler)
•Weighted Fair Queuing (sorted priority scheduler)
WFQ
DRR
•It schedules packets emulating the
behavior of an ideal fluid system (GPS)
•High performance in terms of end-toend and delay jitter
•It provides a fair distribution of service
and a good isolation between flows
•Logarithmic complexity with respect to
the number of flow
•It visits, in a round robin fashion, all nonempty traffic queues: at each turn it sends a
mean amount of data of the flow (quantum)
•It may introduce a higher latency than WFQ
•Computational complexity independent
from the number of queues
•Our goal is to configure DRR weights in order to approximate the
performance of WFQ system in terms of end-to-end delay and delay jitter
3
Reference DiffServ Network Scenario
Expedited
Forwarding
sources
Assured
Forwarding
sources
EF traffic
collectors
Scheduler
Backbone
Link
AF traffic
collectors
Primary Path
100Mbps Links
Best Effort
traffic
1Gbps Link
• We consider a simple DiffServ Model with only three classes
(EF, AF e Best Effort)
• The EF class deliver packets for real-time and delay sensitive
applications
• The AF class carries traffic for applications with less stringent
timing requirements than EF: AF packets should be delivered
within a predefined time interval with low losses.
• The Best Effort applications tolerate with highly variable
transmission delay and delay variation
4
BE traffic
collectors
Traffic characterization with Token Bucket Model
In this study we characterize the AF and EF traffic aggregated flows
through a token bucket model:
~EF
token
rate
 EF
Token
buffer
~EF
token
depth
 EF
s
EF traffic
aggregate
EF class burstiness.
Maximum deviation from
mean long term behavior
output link
AEF (t0 , t )  ~EF  (t  t0 )   EF
5
Mean bitrate of EF
aggregated traffic
Bound on amount of EF traffic
injected into the network during
the interval (t0,t]
Latency-Rate scheduler model
The LR scheduler model is based on the concept of latency and
mean guaranteed rate:
•The latency is the time needed to the LR-scheduler to provide the mean
guaranteed rate to the i-th flow
•The Deficit Weighted round robin scheduler is a LR-scheduler, whose
latency is expressed by the following expression:
 EF  (n  1 
DRR

j  EF
wj
Lmax
)
wEF
r


j  EF
wj QEF

wEF
r
where
wEF
6
EF session weight
n
number of sessions
r
output link capacity
Lmax
maximum packet size
for active sessions
QEF
EF class quantum
Bound on EF class end-to-end delay
The worst-case delay of EF class packets in a network made of a cascade
of k LR-scheduler is given by:
Minimum guaranteed
rate for EF class
 EF
EF
 EF min
Dmax

 EF min
j
EF

Latency of j-th scheduler
for EF class
k
j
j
(  EF
)
  EF
j 1
 EF Burst-size of tokenbucket model for EF
class.
We evaluate the IPDT bound of AF and EF class for the reference DiffServ
network scenario considering the delay constraints
Then, normalizing the weight through AF=wAF/wBE and BE=wEF/wBE , we obtain
a function expressing the EF and AF classes worst-case delay as a function of
TB parameters and quantum
EF
Dmax
( AF ,  BE , EF , QBE )  (
7
1
 BE


NLmax

NQBE 2 NLmax  EF
)( EF 
)  (1  BE )


 AF
r
r
 AF
r
r
r
1
Choice of working parameters
The previous analysis has determined the parameters characterizing the
delay bound. In order to select a configuration of weights we can exploit the
degree of freedom
EF
Dmax
( AF , BE , EF , QBE )
•The ratio AF between AF and BE class quantum is obtained by enforcing a maximum
delay on AF class packets
•By choosing EF on the knee-point of token-bucket curve EF(EFmin), we can have a
tradeoff between the maximum EF class delay and bandwidth requirements
•In order to evaluate the impact BE
quantum on DRR and WFQ performance we
study the behavior of scheduling system in a
limited range of values, observing just small
variations
AEF (t0 , t )   EF min  (t  t0 )   EF
 EF  300KB
 EF  10.3MB
8
Strategy of DRR Weight Configuration
DRR-bnd 240Kb
DRR-bnd 120Kb
DRR-bnd 60Kb
End-to-end delay bound comparison for EF class
DWRR and WFQ by varying the BE quantum
The minimum is obtained by deriving maximum
delay function
EF
Dmax
( AF , BE , EF , QBE )
  NLmax NQBE
r
  EF

0
2
 BE

 BE
AF
WFQ-bnd
 BE 
 AF ( EF  NLmax )
NQBE
Applying this condition to weights associated to DRR to EF, AF e
BE service classes means:
Analytically: The minimization of worst-case delay IPTD EF class
Experimentally: the minimization of performance gap between DWRR and
WFQ in terms of maximum delay and delay variation
9
Simulation Setup
NS-2 simulation topology
Expedited
Forwarding
sources
Assured
Forwarding
sources
EF traffic
collectors
Scheduler
Backbone
Link
Primary Path
100Mbps Links
Best Effort
traffic
1Gbps Link
BE traffic
collectors
Performance Metrics
•IP Transfer Delay (IPTD): end-to-end delay experienced by i-th packet
•IP Delay Variation (IPDV): end-to-end delay variation experienced by
packet with respect to a reference delay
•We evaluate the mean of maximum IPTD and mean IPDV in a set of
five simulations of about 60sec for each BE value
10
.
Simulation Results (maximum IPTD)
Maximum IPTD comparison for EF class (QBE =7.5KB and QBE =30KB)
The worst-case bound is very conservative with respect to results of simulations
but the behavior of both curve is very similar
QBE  7.5KB
 BE  7.38
DRR-bnd
WFQ-bnd
DRR-sim
WFQ-sim
QBE  30 KB
DRR-bnd
WFQ-bnd
DRR-sim
WFQ-sim
 BE  4.47
By assigning to DWRR classes the BE obtained through previous analysis,
we can observe …
•The minimization of worst-case IPDT for EF class packets
•The reduction of loosing of performance between DWRR and WFQ schedulers
11
Simulation Results (average IPDV)
Average IPDV comparisons for EF class between DWRR and WFQ (QBE =7.5KB and 30KB)
QBE  7.5KB
 BE  7.38
DRR-sim
WFQ-sim
QBE  30 KB
DRR-sim
WFQ-sim
 BE  4.47
•Larger the BE Quantum larger the size of DRR frame for a single round-robin
service cycle
•For a large DWRR frame, the inter-departure time of packets delivered in
consecutive rounds may be considerable. Then, it is necessary to avoid the
use of too large BE quantum
12
Second set of simulations
Aggregated
traffic flow
First
simulation
second
simulation
average
bitrate
71.92 Mbps
129 Mbps
peak-rate
0.1sec
interval
98.2 Mbps
308 Mbps
• We incremented the AF class load in terms of mean bitrate and
burstiness, while keeping the same traffic in EF and BE classes
•The AF traffic aggregate flow was obtained by multiplexing of sixty
VIC flows
13
Test results comparisons (worst-case IPTD)
Maximum IPTD comparison for EF class between first and second test
DRR-sim test 1
WFQ-sim test 1
DRR-sim test 2
WFQ-sim test 2
QBE  30 KB
QBE  30 KB
 BE  4.38
• As we could expect, the worst-case IPDT increasing is larger in the case of
DWRR scheduler than WFQ scheduler.
•Since the WFQ scheduler behavior is close to ideal GPS system, it guarantees
a quite perfect flow isolation
•However, for the selected configuration of weights, we reach again the
minimization of DWRR end-to-end transmission delay and the reduction of
performance gap with respect to WFQ
14
Conclusions
This work has led to the definition of an optimization strategy to configure
the bandwidth allocated to different DiffServ flows
Simulation results validate the effectiveness of technique in selecting the
best DWRR operating point
This procedure allows the minimization of worst-case IPDT of privileged
class, while limiting the delay of other classes to prearranged thresold
Moreover, this strategy allow to reduce the differnce in performance
between DRR and WFQ schedulers in terms both of IPDT and IPDV
15