tomography-based overlay network monitoring yan chen, david bindel , randy h. katz

1
Tomography-based Overlay Network Monitoring Yan Chen, David Bindel, Randy H. Katz EECS Department, University of California at Berkeley {yanchen,dbindel,randy}@EECS.Berkeley.EDU Motivation A pplications of end-to-end distance m onitoring O verlay routing/location - P2P system s VPN m anagem ent/provisioning Requirem ents for E2E m onitoring system S calable & efficient:sm allam ount of probing traffic A ccurate:capture congestion/failures S tatic estim ation:G lobalN etw ork Positioning (G N P) D ynam ic m onitoring Loss rates:RO N ( n 2 m easurem ent) Latency:ID M aps,D ynam ic D istance M aps,Isobar Latency sim ilarity under norm alconditions doesn’t im ply sim ilar losses ! N etw ork tom ography Inferring the characteristics of links rather than E2E paths Lim ited m easurem ents -> under-constrained system Existing W ork 1 Path M atrix and Path S pace Path loss rate p ,link loss rate l • Totally s links,path vector v 1 . . ) 1 ( 1 j v t s j j l p x v x v l v p T s j j j s j j j 1 1 ) 1 log( ) 1 log( 1 , } 1 , 0 { , r s r b G b Gx A p B l j Path m atrix G Put all r = O ( n 2 ) paths together Path loss rate vector b 3 Problem Form ulation Given n end hosts on an overlay netw ork and O ( n2 ) paths,how to select a m inim alsubset of paths to m onitor so that the loss rates/latency of all other paths can be inferred. Key idea:select a basis set of k paths that com pletely describe allO ( n 2 )paths ( k «O ( n 2 )) S elect and m onitor k linearly independent paths to com pute the loss rates of basis set Infer the loss rates of allother paths A pplicable for any additive m etrics,like latency End hosts O verlay N etwork O peration C enter 2 S am ple Path M atrix x 1 - x2 unknow n => set of vectors form nullspace T o separate identifiable vs. unidentifiable com ponents: x = xG + xN A llE2E paths are in path space,i.e., GxN = 0 0 1 1 2 ) ( 2 / 2 / 1 0 0 0 1 1 2 ) ( 2 1 2 1 1 3 2 1 x x x b b b x x x x N G G N G Gx Gx Gx Gx b 1 1 1 1 0 0 0 1 1 G 3 2 1 3 2 1 b b b x x x G A D C B 1 2 3 b 1 b 2 b 3 (1,-1,0) link 2 link 1 link 3 (1,1,0) row space (m easured) null space (unmeasured) T ] 0 1 1 [ 4 Algorithm s • Select k = rank( G)linearly independent paths to m onitor U se rank revealing decom position,e.g.,Q R with colum n pivoting Leverage sparse m atrix:tim e O ( rk 2 )and m em ory O ( k 2 ) E.g.,10 m inutes for n = 350 ( r = 61075)and k = 2958 Com pute the loss rates of other paths Tim e O ( k 2 ) and m em ory O ( k 2 ) T opology m easurem ent errors tolerance Care about path loss rates than any interior links Router aliases => assign sim ilar loss rates to allthe alias links Incom plete route => add direct virtuallinks T opology changes A dd/rem ove/change one path incurs O ( k 2 )tim e G G G Gx b b x G x then , with Solve 6 H ow M uch M easurem ent S aved ? Internet has m oderate hierarchicalstructure [T G J+02] If a pure hierarchicalstructure (tree): k = O ( n ) If no hierarchy at all(w orst case,clique): k = O ( n2 ) Internet should fallin betw een … For reasonably large n , (e.g., 100), k = O ( n log n ) B R ITE 20K -node hierarchical topology M ercator284K-node real routertopology 7 Evaluation 8 1 A ustralia 2 Canada 1 H ong K ong 1 Taiw an A sia (2) 2 UK 1 G erm any 1 Denm ark 1 Sw eden 1 France Europe (6) Interna- tional (11) 1 .us 1 .gov 2 .net 3 .org 33 .edu U S (40) # of hosts A reasand D om ains(51 hosts) Extensive sim ulation and PlanetLab experim ents Sim ultaneous loss rate m easurem ents through U D P stream ing Sim ultaneous traceroute A veragely,248 out of 2550 paths have no or incom plete routing inform ation N o router aliases resolved Experim ents:6/24 – 6/27 100 experim ents @ peak hours O n average k = 872 out of 2550 paths Loss rate distribution 25.6% 4.3% 23.9% 31.0% 15.2% 95.9% % 1.0 [0.5, 1.0) [0.3, 0.5) [0.1, 0.3) [0.05, 0.1) lossy path [0.05, 1.0](4.1% ) [0, 0.05) loss rate • A ccuracy A bsolute error • A verage 0.0027 for all paths,0.0058 for lossy paths Relative error factor CD F of 95 percentile for each experim ent 9 9 Intuition through Topology Virtualization Virtual links : minimalpath segm ents w hose loss rates uniquely identified Can fully describe all paths x G : similar form s as virtuallinks 1 2 1 2 3 1’ R eal links (solid)and all ofthe overlay paths (dotted)traversing them V irtualization Virtual links 1’ 2’ 1 2 3 1’ 2’ 4 Rank(G )=1 Rank(G )=2 Rank(G )=3 3’ 4’ 1 1 2 1 2 3 1 1 G 1 0 0 1 1 1 G 1 1 0 0 1 0 1 0 0 1 0 1 0 0 1 1 G A tomography-based overlay network monitoring system - Given n end hosts, characterize O(n 2 ) paths with a basis set of O(nlogn) paths - Selectively monitor O(nlogn) paths to compute the loss rates of the basis set, then infer the loss rates of all other paths • Both simulation and PlanetLab experiment results promising: avg absolute error 0.0027 Conclusi ons Work in Progress • Network diagnostics • Provide it as a continuous service on PlanetLab • More efficient numerical iterative methods for path selection http://www.cs.berkeley.edu/~yanchen/research/wnmms

Upload: zephania-branch

Post on 31-Dec-2015

14 views

Category:

Documents


0 download

DESCRIPTION

Tomography-based Overlay Network Monitoring Yan Chen, David Bindel , Randy H. Katz EECS Department, University of California at Berkeley {yanchen,dbindel,randy}@EECS.Berkeley.EDU. http://www.cs.berkeley.edu/~yanchen/research/wnmms. Conclusions. Work in Progress. - PowerPoint PPT Presentation

TRANSCRIPT

  • Tomography-based Overlay Network MonitoringYan Chen, David Bindel, Randy H. KatzEECS Department, University of California at Berkeley {yanchen,dbindel,randy}@EECS.Berkeley.EDU A tomography-based overlay network monitoring system Given n end hosts, characterize O(n2) paths with a basis set of O(nlogn) paths Selectively monitor O(nlogn) paths to compute the loss rates of the basis set, then infer the loss rates of all other paths Both simulation and PlanetLab experiment results promising: avg absolute error 0.0027ConclusionsWork in Progress Network diagnostics Provide it as a continuous service on PlanetLab More efficient numerical iterative methods for path selectionhttp://www.cs.berkeley.edu/~yanchen/research/wnmms

  • Motivation

    Applications of end-to-end distance monitoringOverlay routing/location - P2P systems VPN management/provisioning Requirements for E2E monitoring systemScalable & efficient: small amount of probing trafficAccurate: capture congestion/failures

    Static estimation: Global Network Positioning (GNP)Dynamic monitoringLoss rates: RON (n2 measurement)Latency: IDMaps, Dynamic Distance Maps, IsobarLatency similarity under normal conditions doesnt imply similar losses !Network tomographyInferring the characteristics of links rather than E2E pathsLimited measurements -> under-constrained system

    Existing Work

    1

  • Path Matrix and Path Space

    Path loss rate p, link loss rate l

    Totally s links, path vector v

    A

    p

    B

    lj

    Path matrix GPut all r = O(n2) paths togetherPath loss rate vector b

    3

  • Problem Formulation

    Given n end hosts on an overlay network and O(n2) paths, how to select a minimal subset of paths to monitor so that the loss rates/latency of all other paths can be inferred.

    Key idea: select a basis set of k paths that completely describe all O(n2) paths (k O(n2)) Select and monitor k linearly independent paths to compute the loss rates of basis setInfer the loss rates of all other pathsApplicable for any additive metrics, like latency

    2

  • Sample Path Matrix

    x1 - x2 unknown => set of vectorsform null spaceTo separate identifiable vs. unidentifiable components: x = xG + xN

    All E2E paths are in path space, i.e., GxN = 0

    4

  • Algorithms

    Select k = rank(G) linearly independent paths to monitorUse rank revealing decomposition, e.g., QR with column pivotingLeverage sparse matrix: time O(rk2) and memory O(k2)E.g., 10 minutes for n = 350 (r = 61075) and k = 2958Compute the loss rates of other paths

    Time O(k2) and memory O(k2)

    Topology measurement errors toleranceCare about path loss rates than any interior linksRouter aliases => assign similar loss rates to all the alias linksIncomplete route => add direct virtual linksTopology changesAdd/remove/change one path incurs O(k2) time

    6

  • How Much Measurement Saved ?

    Internet has moderate hierarchical structure [TGJ+02]If a pure hierarchical structure (tree): k = O(n)If no hierarchy at all (worst case, clique): k = O(n2)Internet should fall in between

    For reasonably large n, (e.g., 100), k = O(nlogn)

    7

  • Evaluation

    8

    Extensive simulation and PlanetLab experimentsSimultaneous loss rate measurements through UDP streamingSimultaneous tracerouteAveragely, 248 out of 2550 paths have no or incomplete routing informationNo router aliases resolvedExperiments: 6/24 6/27100 experiments @ peak hoursOn average k = 872 out of 2550 paths

    Areas and Domains (51 hosts)# of hosts

    US (40).edu33

    .org3

    .net2

    .gov1

    .us1

    Interna-tional (11)Europe (6)France1

    Sweden1

    Denmark1

    Germany1

    UK2

    Asia (2)Taiwan1

    Hong Kong1

    Canada2

    Australia1

  • Loss rate distribution

    AccuracyAbsolute errorAverage 0.0027 for all paths, 0.0058 for lossy pathsRelative error factorCDF of 95 percentile for each experiment

    9

    9

    lossrate[0, 0.05)lossy path [0.05, 1.0] (4.1%)

    [0.05, 0.1)[0.1, 0.3)[0.3, 0.5)[0.5, 1.0)1.0

    %95.9%15.2%31.0%23.9%4.3%25.6%

  • Intuition through Topology Virtualization

    Virtual links: minimal path segments whose loss rates uniquely identifiedCan fully describe all pathsxG : similar forms as virtual links

    Weidong thinks arch-e paper size is the right oneLook under page settings to find the size of this doc seems to fit!Use slides as 205% x 205%