1 Capacity Dimensioning Based on Capacity Dimensioning Based on Traffic Measurement in the Traffic Measurement in the Internet Internet Kazumine Matoba(k- [email protected]) Osaka University Shingo Ata (Osaka City Univ.) Masayuki Murata (Osaka Univ.) Now in Fujitsu Laboratories * *
21
Embed
1 Capacity Dimensioning Based on Traffic Measurement in the Internet Kazumine Matoba([email protected]) Osaka University Shingo Ata (Osaka City Univ.)
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
Capacity Dimensioning Based on Traffic Capacity Dimensioning Based on Traffic Measurement in the InternetMeasurement in the Internet
Capacity Dimensioning Based on Traffic Capacity Dimensioning Based on Traffic Measurement in the InternetMeasurement in the Internet
TCP throughput (estimated value)*take no thought of sender side configuration
* J.Padhye, V.Firoiu, D.Toesley.,and J.Kurose, “Modeling TCP throughput: A simple model and its empirical validation”, Proceedings of ACM SIGCOMM’98, pp.303-314, September 1998.
expected window size of TCP connection
actual throughput << estimated throughput
① sender side configuration
② check buffer size (receiver side)
b : duplicate ACK numberp : packet loss rate
2
32
3)1(8
32
][
bb
bpp
bb
WE 2
32
3)1(8
32
][
bb
bpp
bb
WE
buffer size of receiver hostexpected window size
:: ][
maxWWE
maxWE[W ] maxWE[W ]
③ check network congestion (network condition)
congested router queue and/or drop packets
large RTT and high packet loss rate
*
bottleneck is sender!
buffer size is insufficient!
network is congested!
7
- buffer size is insufficient
- when buffer size is increased, TCP throughput becomes high
Bottleneck Classification Process measure RTT, Packet Loss Rate, and TCP throughput
check socket buffer size of receiver host
repeat measurement, changing socket buffer size
- sufficient size of buffer
- high packet loss rate
- large RTT
- no relationship between buffer size and throughput- actual throughput << estimated throughput
① sender side configuration
② buffer size (receiver side)
③ network congestion (network condition)
8
Server 2Server 1
Classification Experimens in the Internet
0
10
20
30
40
8 16 24 32 40 48 56 64
Buffer Size (KByte)
Th
rou
gh
pu
t (M
bp
s)
0
10
20
30
40
8 16 24 32 40 48 56 64
Buffer Size (KByte)
Th
rou
gh
pu
t (M
bp
s)
0
10
20
30
40
8 16 24 32 40 48 56 64
Buffer Size (KByte)
Th
rou
gh
pu
t (M
bp
s)
Estimated throughputActual throughput
No special feature
buffer increase
⇒Throughput increase
Larger than Actual throughput
Buffer size is insufficient
Network Condition
Receiver SideSender Side
05
10152025303540
8 16 24 32 40 48 56 64
Buffer Size (KByte)R
ou
nd
Tri
p T
ime (
ms)
0
0.05
0.1
0.15
0.2
0.25
0.3
Pac
ket
Lo
ss R
ate
(%
)
Packet Loss Rate
Round Trip Time
Large RTT
High Packet Loss Rate
For each bottleneck,
their characteristics can be found
For each bottleneck,
their characteristics can be found
Server 3
9
Contents
Capacity Dimensioning1. Measurement Based Identification Method of
Performance Bottleneck
2. Measurement Method of Bottleneck Link Utilization
3. Capacity Dimensioning Based on Traffic Measurement
Conclusion
10
Estimate Link Utilization
measure and estimate capacity of bottleneck linkPathchar, Pchar, Clink, etc.
ccrr
bb
rrbb
uu cc
estimate bottleneck link utilization
measure the amount of cross traffic passing through the bottleneck link
11
need these items!
Estimate Throughput of Cross Traffic
assume probe packets are queued at router n
cross trafficcr
probe packet
Bottleneck Router n Router n+1
Bottleneck Link
n+1lnl
sb llr nnc 1the amount of
cross traffics :size of probe packet
b :capacity of link n+1
throughput of cross trafficn
nc
l
sb lr
1
client host send ICMP ECHO packet continuously
router n and n+1 return reply packets
observe returned packet and calculate interval of arrival time
client host send ICMP ECHO packet continuously
router n and n+1 return reply packets
observe returned packet and calculate interval of arrival time nnll n+1n+1ll
12
Estimation Results of Link Utilization
measure bottleneck link between two universities
Estimation error is under 10% in link utilization
link utilization
(cross traffic)
estimation
15.9% (0.95 Mbps) 22.8%
27.9% (1.7 Mbps) 27.8%
38.5% (2.3 Mbps) 32.1%
69.5% (4.1 Mbps) 60.3%
77.6% (4.6Mbps) 81.6%
Bottleneck link 6MbpsBottleneck link 6Mbps
Osaka City Univ.Osaka Univ.
Measurement HostMeasurement Host
Router nRouter nRouter n+1Router n+1
Target HostTarget Host
13
Contents
Capacity Dimensioning1. Measurement Based Identification Method of
Performance Bottleneck
2. Measurement Method of Bottleneck Link Utilization
3. Capacity Dimensioning Based on Traffic Measurement
Conclusion
14
Capacity Dimensioning
from user’s viewpointfulfill a demand of user
Designing network for each bottleneck1. sender or receiver side configuration
2. network configuration (link bandwidth)
based on end-to-end network characteristicsbased on end-to-end network characteristics
propose a design framework
15
Capacity Dimensioning 1:Bottleneck Resides in the End Host
Sender side configurationremove the cause of bottleneck
• Ex. configure rate shaping, upgrade hardwares, etc.
Receiver side configuration (buffer size)increase buffer size to ff
RTTRTTttff '''' tt '': throughput which user needs: throughput which user needs
: Round Trip Time: Round Trip TimeRTTRTT
''
Bottleneck may shift to another areaBottleneck may shift to another area
need to upgrade link capacity?
16
Designing Link Capacity:Bottleneck Resides in the End Host
most thin available bandwidth link
timetime
cross trafficcross traffic
user’s trafficuser’s trafficCC
ttCCAA
tt
CC))11((
link utilization
tt
ttttAACC ''
lower bound of new link capacity
lower bound of new link capacity
link capacitylink capacity
determine new link capacity
C : link capacity
A : throughput of cross traffict : throughput of user’s traffic
17
user traffic
both of the user and cross traffic
Capacity Dimensioning 2:Bottleneck Resides in Network
timetime
send data
timetime
cross trafficcross traffic
user’s traffic tuser’s traffic t
timetime
cross trafficcross traffic
timetime
cross trafficcross traffic
user’s trafficuser’s trafficsend data
cross traffic Across traffic Alink capacitylink capacity
link capacitylink capacity
same as the former
capacity decision
ttAC '
predict amount of future cross traffic
upgrade
cross traffic may increase
cross traffic may increase
18
Designing Link Capacity:Bottleneck Resides in Network
assume and simplify the situationbottleneck router is M/M/1 queueing system
current link capacity C will increase to C’ =a C
choose following user traffic will increase to t’
aaTTbpbptt
TT
2233
''11
T : packet service time at bottleneck router
b : the number of arrival packets notified one ACK
p : packet loss rate
Ta : constant (=RTT-T)
T : packet service time at bottleneck router
b : the number of arrival packets notified one ACK
p : packet loss rate
Ta : constant (=RTT-T)
19
Conclusion
Capacity dimensioning based on traffic measurement in the Internet
Identification method of performance bottleneck• classify into 3 kinds of bottlenecks
Measurement method of link utilization• focused on the bottleneck link
Capacity dimensioning• Modeling the router as an M/M/1 queueing system
20
The End
21
Available BandwidthAvailable Bandwidth
Existing Method to EstimateLink Utilization
Bottleneck linkBottleneck link
packet pair
packet stream
probe packet affects measurement resultbut not take into consideration