Transcript WF2Q with Maximum Rate Control

On Maximum Rate Control of
Weighted Fair Scheduling
Jeng Farn Lee
1
Outline





Introduction
Related Work
WF2Q with maximum rate control
Simulations
Conclusions
2
Introduction


Current service disciplines provided
minimum performance guarantees, but not
maximum rate constraint
Max-Rate Control is needed


Control lease line’s maximum services rate
Restrict specific applications’ total traffics to
enforce some management policies
3
Introduction (cont’d)


Ban over-provisioning in a link-sharing
environment (e.g. WF2Q)
Stabilize the throughput to smooth media
streaming in order not to overflow receiving
buffers or cause packet drop
4
GPS

GPS (Generalized Processor Sharing)

A fluid system
traffic is infinitely divisible
 all the traffic streams can receive service
simultaneously


Each session i is assigned a fixed real-valued
positive parameter  i
5
GPS (cont’d)
session is
idle after time 10
6
Virtual Clock


Implementation of PGPS
Virtual clock is a clock to keep a
normalized time as a standard reference for
all sessions/packets.
V (0)  0

V (t j 1   )  V (t j 1 ) 
 i
iB j
7
Two-Stage Rate-Control Service
Model
one regulator for each of
the N sessions
Regulated Traffic
Regulator 1
Regulator 2
Output
Iutput
Regulator N
Rate Controller
Scheduler
8
Two-Stage Rate-Control Service
Model (cont’d)

Drawbacks

When move packets from regulator queue to eligible
queue

Timer




the system must use one interrupt to change the status per packet
Time-framing (system accuracy v.s. time granularity)
Event-Driven (high uncertainty)
It still needs to modify the scheduling algorithm to
distribute the excess bandwidth to other sessions
9
Policer-Based Rate-Control
Service Model
conforming
packets
Policer
Output
Input
Scheduler
10
Policer-Based Rate-Control
Service Model (cont’d)

Drawbacks

Token bucket


Token buffer allows traffic exceed the maximum
rate
Leaky bucket

Not allow traffic burst
11
Simulation environment

ns2




Version : ns-allinone2.1b6
WFQ patch 1.1a1
We implement of policer-based rate-control
service model and WF2Q-M
topology
S1
n1
n2
R1
10Mbps,
2ms
12
Traffic pattern



UDP Exponential ON/OFF traffic
The packet size of ON period : exponential
distribution with mean (1000, 950 and 900 bytes)
The maximum rate of the session is 4Mbps
13
Traffic pattern
6
5
4
3
2
1
0
0
10
20
30
40
50
60
70
14
Token bucket with r = 4Mbps,
B=0.25Mb
Loss rate : 0.211%
Over max rate rate : 12.96%
6
5
4
3
2
1
0
0
10
20
30
40
50
60
70
15
Leaky bucket r=4Mbps
Loss rate : 58.89%
2
1.5
1
0.5
0
0
10
20
30
40
50
60
70
16
Wf2q-m buffer size 0.25Mb
Loss rate : 0.219%
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
0
10
20
30
40
50
60
70
17
18
GPS-M


An extension of GPS
A session can be “normal” session or “maximum
rate constrained” session. If a maximum rate
constrained has shared bandwidth greater than the
maximum rate,



It receives the maximum rate;
GPS-M distributes the excess bandwidth to others
weightily
WF2Q-M use the same link sharing principle as
GPS-M
19
GPS-M
Resource Allocation
ex. 10 packets per second, reserved bandwidth
5:2.5:1.25:1.25
GPS and GPS-M
GPS
5
GPS-M
2.5
5
4
1.25
2.5
3
1.25
2.5
3
Max Rate=4
20
Features of WF2Q-M




Merge packet eligible time into virtual starting
time
Only the packets have started receiving service in
GPS-M can be selected for transmission
Adjust the ticking rate of the system virtual clock
to distribute the excess bandwidth from saturated
queues to other sessions
Use the same real clock/virtual clock ratio to
transfer real clock for packets of saturated queues
to virtual clock
21
Virtual Clock Adjustment
V (0)  0
V (t   )  V (t ) 

ratio(t )

jB ( )
ratio(t)=

jB ( )
j *
j


kB p ( )

jB ( )
C
k
*C
j
P
kB p ( )
k
22
Marge eligible time into virtual
starting time
The virtual starting and finishing times of packets of Bp(p)
k 1
k
k k 1 Li
ei  max(ai , ei 
)
Pi
Sik  max{V (eik ), Fik 1}}
k
L
Fik  Sik  i / ratio(t )
Pi
Sik  max{V (aik ), Fik 1}
k
L
Fik  Sik  i / ratio(t )
Pi
23
WF2Q-M: Virtual Times
for Bp (p)
Sik  max{V (aik ), Fik 1}
k
L
Fik  Sik  i / ratio(t )
Pi
others
Sik  max{V (aik ), Fik 1}
k
L
Fik  Sik  i
i * C
24
Simulations
ns2




Version : ns-allinone2.1b6
WFQ patch 1.1a1
WF2Q and WF2Q-M
topology
S1
n1
n2
10Mbps,
2ms
S2
R1
S3
S4
n3
10Mbps,
2ms
R2
n4
10Mbps,
2ms
R3
R4
25
Simulations (cont’d)

Data sending rate : 5Mbps



Packet size : Uniform(100,1500) bytes
Data type : UDP
Maximum rate of session 3 is 3Mbps
26
Simulation Result (WF2Q)
S1 reserved bandwidth 10%
S2 reserved bandwidth 15%
S3 reserved bandwidth 25%
S4 reserved bandwidth 50%
6
throughput (Mbps)
5
4
3
2
1
0
0
2
4
6
8
10
12
14
time (sec)
27
Simulation Result (WF2Q-M)
S1 reserved bandwidth 10%
S2 reserved bandwidth 15%
S3 reserved bandwidth 25%
S4 reserved bandwidth 50%
6
Throughput (Mbps)
5
Maximum rate is
3Mbps
4
3
2
1
0
0
2
4
6
8
10
12
14
time (sec)
28
Conclusions

we propose a new service discipline WF2Q-M





guarantee minimum service rates as WF2Q
provide maximum service rate constraint
merge packet eligible time into its virtual starting time
to reduce complexity
virtual clock adjustment allows the sharing of excess
bandwidth to non saturated sessions
WF2Q-M performance is bounded by a fluid reference
mode
29
Thank You!
Jeng Farn Lee
[email protected]
30