Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Bufferbloat Detection In Network Paths
Danny Groenewegen1, Harald Kleppe1
1System and Network EngineeringUniversiteit van Amsterdam
Michiel Leenaars22NLnet Foundation
30th of June 2011RP2 presentations
BufferbloatDark buffers in the Internet
Bufferbloat is the existence of excessively large (bloated) buffers insystems, particularly network communication systems.
I Every network protocol is affectedI Many of them rely on timely arrival of packets
I The problem was already discussed in RFC 970
I Cheaper memory, increasing buffers
I Increasing traffic
Detecting bufferbloatResearch question
Is it possible to determine which device in a path causesbufferbloat?
I Is it possible to quantify the effects of bufferbloat?
I Can the same methods be applied on the Internet, where theyare influenced by other unpredictable traffic?
Measuring link characteristics with pathcharProbing each hop in a path
I Vary TTL to probe each node in a path
I Difference between n − 1 and n gives link characteristics
1
1Allen Downey. Using pathchar to estimate Internet link characteristics;ACM SIGCOMM, pages 241-250, 1999
Measuring link characteristics with pathcharEstimate bandwidth and latency
I Varying packet sizes givesbandwidth and latency
I Multiple repetitions givequeuing and loss information
Estimating queue delaysFrom pathchar to measuring bufferbloat
I Improve the accuracy of queue delay estimatesI Probing each hop at (almost) the same timeI Only probing with one packet size
I Observing the minimum RTT in a congested pathI Bufferbloat effects are not always visible
I Measure with five minute intervals
Quantifying bufferbloatBufferbloat scoring model
gr ad e = 0
f o r each measurement i n t e r v a l :median = median o f queue d e l a y sgr ad e += ( median / 0 . 0 1 0 ) / number o f i n t e r v a l s
i f g ra de > 1 0 :g r ad e = ”10+”
Tests and results
I Simulated experiments in a lab setup
I Experiments on the Internet
I Validate the toolI Detect queue distribution in a pathI Compare bufferbloat scoring
ResultsLab test #1
I 10 and 100 Mbit bottlenecksI Saturate bottlenecksI Identify second bottleneck behind an intermediate bottleneck
Router #1 Router #2 Router #3
Router #4Router #5
Host #1
Host #2Router #6
ResultsLab test #1
I 10 and 100 Mbit bottlenecksI Saturate bottlenecksI Identify second bottleneck behind an intermediate bottleneck
Router #1 Router #2 Router #3
Router #4Router #5
Host #1
Host #2Router #6
ResultsTest setup; Link 3: 10Mb/s; Link 6: 100Mb/s; Bufferbloat score: 10+
0
100
200
300
400
500
600
700
800
900
0 10 20 30 40 50
Add
ed Q
ueue
del
ay(m
s)
Time (minutes)
Queue delay on top of minimum path RTT - bufferbloat-testsetup2
Minimum Path RTTLink 1Link 2
Link 3Link 4Link 5
Link 6
ResultsConsumer ISP connection; Bufferbloat score: 5.0
I Probing from Diemen to Norway
I Saturating traffic from OS3 lab
I 25 / 1.5 Mbit consumer grade connection
ResultsConsumer ISP connection
0
100
200
300
400
500
600
Jun 29 17:45
Jun 29 19:10
Jun 29 20:25
Add
ed Q
ueue
del
ay(m
s)
Time
Queue delay on top of minimum path RTT - bufferbloat-diemenuio
Min pathLink 1
Link 2Link 3
Link 4Link 5
Link 6Link 7
Link 8Link 9
Link 10
ResultsChicago - Tokyo
I From a VPS in Chicago
I Towards a RIPE TTM host in Tokyo
I Eight hop path, trans-pacific
ResultsChicago - Tokyo: Bufferbloat score: 8.6
0
50
100
150
200
250
Jun 29 02:36
Jun 29 04:41
Jun 29 06:41
Jun 29 09:26
Jun 29 10:56
Jun 29 12:46
Add
ed Q
ueue
del
ay(m
s)
Time
Queue delay on top of minimum path RTT - bufferbloat-usatokyov4
Min pathLink 1
Link 2Link 3
Link 4Link 5
Link 6Link 7
Link 8
Results Summary
I Queue delay can be determined per link
I Possible to observe near zero queue delay
I Generally slightly less queue delay on IPv6 links than IPv4
Conclusion
I Possible to detect bufferbloated linkI Determining which queueI Layer 2 components
I Methods seem to work on the InternetI Scoring model needs fine tuning
Questions?
Thanks for listening!
Tool: https://github.com/hkleppe/Buffchar
More info: http://www.bufferbloat.net