インターネット入門

Download Report

Transcript インターネット入門

Dynamic threshold Control of RED for
establishing fairness among thousands of TCP
connections
Go Hasegawa
Osaka University, Japan
[email protected]
8th, Nov 2001
APSITT2001
1/15
Backgrounds

Internet Router Buffer
 Tail-Drop Router
 Simple, easy implementation
 Discard arriving packets when buffer is full



Bursty packet loss
Poor performance

Throughput

fairness among connections
RED (Random Early Detection)
 Discard incoming packets with a
probability
8th, Nov 2001
APSITT2001
certain
2/15
RED (Random Early Detection)
Probability is changed according to average
queue length
Avoid buffer overflow, keep queue length low


Packet Discarding Probability
Queue Length [packets]
80
1
maxth
Queue Length
60
40
maxp
20
minth
maxth
Average Queue Length
minth
0
100
Average Queue Length [packets]
8th, Nov 2001
APSITT2001
105
110
Time [sec]
115
120
3/15
Researches on RED

Achieve better performance than
Tail-Drop
 Throughput, Fairness
among
connections

Difficult to set control parameters
 Depends
on network condition, # of
active connections, …
8th, Nov 2001
APSITT2001
4/15
Researches on RED (2)

Many enhanced algorithm have
been proposed
 Introducing
new control parameters
 Parameter setting problem remains
Focus only on throughput
 Small number of connections

8th, Nov 2001
APSITT2001
5/15
Objectives

Evaluate fairness of RED with many
connections
 Compare with

Tail-Drop
Propose new algorithm of RED
 Set
threshold values dynamically
 Easy parameter setting
 Provide good fairness
8th, Nov 2001
APSITT2001
6/15
Network Model
Router Buffer Size:
B packets
Sender Host 1
1.5Mbps, 4msec
100Mbps, 2msec
Sender Host 2


Sender Host N

Receiver Host
N sender hosts transmit infinite size of
data to receiver host by TCP Reno
Evaluate fairness among mean
throughputs of N TCP connections
Measure: Fairness Index

x

f 
N
2
i 1 i
N
2
i 1 i
N x

8th, Nov 2001
0 (unfair) < f < 1 (completely fair)
APSITT2001
7/15
Evaluation Result (1): B=1000 [packets]
1
Fairness Index
0.8
0.6
0.4
0.2
0



8th, Nov 2001
1
TailDrop
RED maxth = 15
RED maxth = 800
10
100
1000
N: Number of TCP connections
Tail-Drop shows degrades fairness when N is small
Recommended parameter set of RED shows bad fairness
especially when N is large
RED with appropriate parameters shows good fairness
regardless of N
APSITT2001
8/15
Evaluation Result (2): B=10000 [packets]
1
Fairness Index
0.8
0.6
0.4
0.2
0


8th, Nov 2001
TailDrop
RED maxth = 15
RED maxth = 2000
1
10
100
1000
N: Number of TCP connections
Tail-Drop and RED with recommended parameter set
cannot provide fairness
Best setting of RED changes according to network
condition
APSITT2001
9/15
Fairness of RED

Recommended parameter set cannot provide
fairness when N is large



Threshold value (maxth) should be set carefully
to keep fairness
Appropriate value will be changed by various
factors


8th, Nov 2001
When average queue length becomes maxth, RED
discards incoming packets in burst as Tail-Drop
Bandwidth, Buffer size, # of connections, delay, …
Static parameter setting is impossible
APSITT2001
10/15
Proposed Algorithm: dt-RED
Queue Length [packets]


4000
3000

2000
1000
maxth
Queue Length
Avg. QL
0
100
200
300
Time [sec]
8th, Nov 2001

400
500
Observe average queue length
at regular intervals
Change threshold values (maxth,
minth) according to the average
queue length
Keep average queue length
between maxth and minth
Avoid bursty packet losses at
the RED router buffer
APSITT2001
11/15
Evaluation Result (3): B=1000 [packets]
1
Fairness Index
0.8
0.6
0.4
TailDrop
RED
max
th = 15
0.2
RED maxth = 800
dt-RED
0
1
10
100
1000
Number of TCP connections

8th, Nov 2001
dt-RED can provide the best fairness
among all algorithms
APSITT2001
12/15
Evaluation Result (4): B=1000 [packets]
1
Fairness Index
0.8
0.6
0.4
0.2
0

8th, Nov 2001
TailDrop
RED maxth=15
RED maxth=2000
dt-RED
1
10
100
1000
Number of TCP connections
dt-RED shows good fairness regardless of the
number of connections and buffer size
APSITT2001
13/15
Characteristics of dt-RED

Always provide the same fairness as
original RED with best parameter set
 Automatically
appropriately

set threshold values
Show good performance with one
parameter set, regardless of network
condition
 Remove the
difficulty of parameter setting in
original RED
8th, Nov 2001
APSITT2001
14/15
Conclusion
Fairness of RED algorithm is largely
dependent on control parameters
 Static parameter setting is difficult
 dt-RED can set parameters
appropriately according to dynamic
change of network condition

8th, Nov 2001
APSITT2001
15/15