On the Impact of Clustering on Measurement Reduction May 14 th, D. Saucez, B. Donnet, O. Bonaventure Thanks to P. François Université catholique de Louvain
Measurements to Improve netapps/service performance Bandwidth? Delay? Loss?
3 ? ? ? ? ? ? ? ? ? ? ? ? Scalability issues with large-scale measurements
4 How to reduce the measurement overhead? Limit the number of measured destinations Clustering Limit the number of measuring sources Collaboration
5 Limit the number of measured destinations Group destinations into Clusters
6 Clustering techniques Geographic Clustering Group nodes by city n-agnostic clustering [1] group nodes by /n prefix AS Clustering [2] group nodes by Autonomous System BGP Clustering [3] group nodes by longest match BGP prefix [1] Szymaniak, M. et al., Practical large-scale latency estimation. Computer Networks, 2008 [2] Krishnamurthy, B., Wang, J., Topology modeling via cluster graphs. ACM SIGCOMM Workshop on Internet Measurement (IMW), 2001 [3 ]Krishnamurthy, B., Wang, J., On network-aware clustering of web clients. ACM SIGCOMM, 2000
7 How clustering impacts the accuracy?
8 Evaluation setup Maxmind + Routeviews 1month traceroute traces (Archipelago) Two monitors: san-us (San Diego, US) bcn-es (Barcelona, SP)*
9 RTT error (bcn-es) Geographic, AS n-agnostic, BGP 15% with more than 100% error 10% with more than 200% error 90% with less than 50% error 50% with less than 10% error
10 Clustering reduces the number of measured destinations without loosing too much accuracy can we reduce the number of source of measurements?
11 Limit the number of measuring sources Make measurement sources collaborating
12 Collaboration fundamentals Popular destinations are measured by several nodes Popularity d : #nodes measuring d Different collaboration approaches Centralized authority/measurement source Distributed measurements (ICS)
13 How much reduction can we obtain?
14 When can we observe measurement reduction? Clustering reduces measurements if a cluster C covers at least two measured destinations Collaboration reduces measurements if at least two topologically closed sources have to measure the same destination
15 Evaluation setup Campus traffic UCL, 1 link to 1 month full NetFlow traces 7.45 TB of filtered outgoing traffic 10K sources, 36M destinations
16 Will collaboration help? 74% of the destinations are contacted by only 1 source Some destinations are contacted by 1K+ sources! Few percents are contacted by 10+ sources
17 Will clustering help? At least 45% of the clusters cover more than 10 nodes # of destinations
18 Conclusion Clustering/Collaboration to reduce measurement overhead Reduction/accuracy tradeoff Simple, though efficient techniques, tend to preserve accuracy
19 Questions?
20 Backup
21 Combine Clustering and Collaboration
22 Hop error (bcn-es) 0% more than 50% error 10% more than 50% error bigger the n, smaller the error Geographic, AS n-hybrid, n-agnostic, BGP
23 Error variation inside clusters 75 th percentile 50 ty percentile 25 th percentile
24 The reduction Collaboration only: 40% gain 20-hyb only: 62% gain 20-hyb + Collaboration: 99% gain Collaboration + Clustering always better than clustering or collaboration only
25 Are clustering and collaboration so different? Let C, a cluster of nodes to measure Let S C, the set of nodes measuring C S C is cluster nodes in S C can collaborate => S C is the set of collaborating nodes
n-hybrid Clustering / / / / /24... A B C A B C BGP clusters / /20 20-hybrid clusters BGP prefixes can be huge: => Group nodes by longest match BGP prefix down to a given length
27 traceroute to ( ), 30 hops max, 40 byte packets ( ) ms ms ms 2 c hsd1.ga.comcast.net ( ) ms ms ms 3 ge-2-1-ur01.a2atlanta.ga.atlanta.comcast.net ( ) ms ms ms 4 te-9-1-ur02.a2atlanta.ga.atlanta.comcast.net ( ) ms ms ms 5 te-9-3-ur01.b0atlanta.ga.atlanta.comcast.net ( ) ms ms ms 6 po-4-ar01.b0atlanta.ga.atlanta.comcast.net ( ) ms ms ms 7 pos cr01.atlanta.ga.ibone.comcast.net ( ) ms ms ms 8 te-9-1.car1.Atlanta2.Level3.net ( ) ms ms ms 9 ae ebr2.Atlanta2.Level3.net ( ) ms ms ms 10 ae-3.ebr2.Chicago1.Level3.net ( ) ms ms ms 11 ae car1.Chicago1.Level3.net ( ) ms ae car1.Chicago1.Level3.net ( ) ms ae car1.Chicago1.Level3.net ( ) ms... traceroute to ( ), 30 hops max, 40 byte packets ( ) ms ms ms 2 c hsd1.ga.comcast.net ( ) ms ms ms 3 ge-2-1-ur01.a2atlanta.ga.atlanta.comcast.net ( ) ms ms ms 4 te-9-1-ur02.a2atlanta.ga.atlanta.comcast.net ( ) ms ms ms 5 te-9-3-ur01.b0atlanta.ga.atlanta.comcast.net ( ) ms ms ms 6 po-4-ar01.b0atlanta.ga.atlanta.comcast.net ( ) ms ms ms 7 pos cr01.atlanta.ga.ibone.comcast.net ( ) ms ms ms 8 * * * 9 ae ebr2.Atlanta2.Level3.net ( ) ms ms ms 10 ae ebr3.Atlanta2.Level3.net ( ) ms ae ebr3.Atlanta2.Level3.net ( ) ms ae ebr3.Atlanta2.Level3.net ( ) ms 11 ae-7.ebr3.Dallas1.Level3.net ( ) ms ms * 12 ae-3.ebr2.LosAngeles1.Level3.net ( ) ms ms ms 13 ae csw2.LosAngeles1.Level3.net ( ) ms ae csw1.LosAngeles1.Level3.net ( ) ms ms 14 ge-9-2.core1.LosAngeles1.Level3.net ( ) ms ge-5-2.core1.LosAngeles1.Level3.net ( ) ms ge-5-1.core1.LosAngeles1.Level3.net ( ) ms... Traceroute verdict*
28 N-hybrid example / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / /24 Level 3: /8 ? BGP: /24 20-hybrid: /24 ? BGP: /9 20-hybrid: /20 ? BGP: /9 20-hybrid: /20 BGP (Routeviews) Natural follow up, came for free → dessin
29 References [1] Xie et al., P4P: Provider Portal for Applications, in Proc. ACM SIGCOMM, 2008 [2] Aggarwal et al., Can ISPs and P2P systems co-operate for improved performance?, ACM SIGCOMM Computer Communications Review (CCR), 37(3):29–40, July 2007 [3] Saucez et al., Interdomain Traffic Engineering in a Locator/Identifier Separation Context, Internet Network Management Workshop 2008 [4] Dabek et al., Vivaldi, a decentralized network coordinated system. ACM SIGCOMM, 2004 [5] Krishnamurthy, B., Wang, J., Topology modeling via cluster graphs. ACM SIGCOMM Workshop on Internet Measurement (IMW), 2001 [6] Szymaniak, M. et al., Practical large-scale latency estimation. Computer Networks, 2008 [7 ]Krishnamurthy, B., Wang, J., On network-aware clustering of web clients. ACM SIGCOMM, 2000