1 network telecommunication group university of pisa - information engineering department january 31...
TRANSCRIPT
1
Network Telecommunication Group
University of Pisa - Information Engineering departmentJanuary 31 2005
Speaker:Raffaello Secchi
Authors:Davide AdamiStefano GiordanoMichele PaganoRaffaello Secchi
Optimization of Scheduling Algorithm Parameters in a DiffServ Environment
2
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
3
• In this work we considered two different proportional schemes•Deficit Round Robin (frame-based scheduler)•Weighted Fair Queuing (sorted priority scheduler)
Scheduling Algorithms
•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
s
Server
Output link
W1
W2
W3
The weight associated to the i-th queue is proportional to the percentage of output capacity
The weight associated to the i-th queue is proportional to the percentage of output capacity
WFQ•It schedules packets emulating the behavior of an ideal fluid system (GPS)•High performance in terms of end-to-end 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
DRR•It visits, in a round robin fashion, all non-empty 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
4
Reference DiffServ Network Scenario
AF traffic collectors
Primary Path
Backbone Link
100Mbps Links1Gbps Link
Expedited Forwarding
sources
Assured Forwarding sources
Best Effort traffic
EF traffic collectors
BE traffic collectors
Scheduler
• 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
5
Traffic characterization with Token Bucket Model
In this study we characterize the AF and EF traffic aggregated flows through a token bucket model:
EF class burstiness. Maximum deviation from mean long term behavior
EF
EF~ Mean bitrate of EF aggregated traffic
EFEFEF ttttA )(~),( 00 Bound on amount of EF traffic injected into the network during the interval (t0,t]
Tokenbuffer
s
EF~
EF traffic aggregate
EF
tokenrate
tokendepth
output link
6
Latency-Rate scheduler model
EFj EFj r
Q
w
w
r
L
w
wn EF
EF
j
EF
jDRR
EF
max)1(
EFQ
maxLEFw
n
r
maximum packet size for active sessions
EF class quantum
EF session weight
number of sessions
output link capacity
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:
where
7
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
Latency of j-th scheduler for EF class
Burst-size of token-bucket 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
rr
NL
r
NQ
r
NL
rQD EFBE
AF
BEEF
AFBEBEEFBEAF
EF
maxmax
max2
)1())(11
(),,,(
k
j
jEF
jEF
EF
EFEFD1min
max )(
jEF
EF
minEF
8
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
•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
Choice of working parameters
KBEF 300
MBEF 3.10
EFEFEF ttttA )(),( 0min0
),,,(max BEEFBEAFEF QD
•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
9
DRR-bnd 240KbDRR-bnd 120KbDRR-bnd 60Kb
WFQ-bnd
The minimum is obtained by deriving maximum delay function
0),,,(
2maxmax
AF
BE
BE
EF
BE
BEEFBEAFEF QNNLQD
r
End-to-end delay bound comparison for EF class DWRR and WFQ by varying the BE quantum
BE
EFAFBE QN
NL )( max
Analytically: The minimization of worst-case delay IPTD EF classExperimentally: the minimization of performance gap between DWRR and WFQ in terms of maximum delay and delay variation
Applying this condition to weights associated to DRR to EF, AF e BE service classes means:
Strategy of DRR Weight Configuration
10
Simulation Setup
.
Primary Path
Backbone Link
100Mbps Links1Gbps Link
Expedited Forwarding
sources
Assured Forwarding sources
Best Effort traffic
EF traffic collectors
BE traffic collectors
Scheduler
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
NS-2 simulation topology
11
DRR-bndWFQ-bndDRR-simWFQ-sim
DRR-bndWFQ-bndDRR-simWFQ-sim
KBQBE 5.7 KBQBE 30
47.4BE38.7BE
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
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
Simulation Results (maximum IPTD)
12
Simulation Results (average IPDV)
Average IPDV comparisons for EF class between DWRR and WFQ (QBE =7.5KB and 30KB)
DRR-simWFQ-sim
DRR-simWFQ-simKBQBE 5.7 KBQBE 30
47.4BE38.7BE
•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
13
Second set of simulations
• 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
Aggregated traffic
flow
First simulation
second simulation
average bitrate
71.92 Mbps 129 Mbps
peak-rate 0.1sec interval
98.2 Mbps 308 Mbps
14
Test results comparisons (worst-case IPTD)
Maximum IPTD comparison for EF class between first and second test
DRR-sim test 2WFQ-sim test 2
KBQBE 30
DRR-sim test 1WFQ-sim test 1
KBQBE 30
38.4BE• 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
15
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