Discovering Routers in Load-balanced Paths

Jean-Fran ois Grailet, Benoit Donnet

Universitéde Liège Monteore Institute

Belgium

rstname.lastnameulg.a .be

ABSTRACT

Usually,asetof tra eroutemeasurements olle ted

foralargeamountoftargetaddresses ontainoneor

sev-eralroutehopsatwhi htheIPinterfa esvaryfromone

measurement to another. These variations o ur even

ifseveralmeasurementsshare thesame lengthandthe

samelasthops.Thisislikelya onsequen e ofload

bal-an ing, a tra engineering poli ywhi h aimsat

shar-ing the load between multiple paths to ensure quality

of servi e. In this paper, we onsider the problem of

ondu ting alias resolution on IP interfa esdis overed

via tra eroute and that are involved in load

balan -ing. By ondu ting alias resolution in su h a ontext,

wewant to verify if the IPinterfa es involved in load

balan ing belong to uniquerouters, and more broadly,

how relevant is alias resolution in this ontext. To do

so,weuseaslightlyeditedversion of TreeNET,a

topol-ogydis overy toolwhi h relies on a tree-like stru ture

basedontra eroutemeasurementstomapatarget

do-main.Theupgraded TreeNETalong ourmeasurements

arebothfreelyavailableonline.

CCS CONCEPTS

ˆ Networks  Network measurement; T

opol-ogy analysis and generation;

KEYWORDS

TreeNET,subnettopology,aliasresolution,tra eroute

ACMReferen eFormat:

Jean-Fran oisGrailet,BenoitDonnet.2017.Dis overingRouters

inLoad-balan edPaths. In Pro eedings of ACM CoNEXT

StudentWorkshop,Seoul,Korea,2017(CoNEXT'17),3pages.

https://doi.org/10.1145/nnnnnnn.nnnnnnn

Permissionto makedigital or hard opies ofpart or all of this

workforpersonalor lassroomuseisgrantedwithoutfeeprovided

that opiesarenotmadeordistributedforprotor ommer ial

advantageandthat opiesbearthisnoti eandthefull itationon

therstpage.Copyrightsforthird-party omponentsofthiswork

mustbehonored.Forallotheruses, onta ttheowner/author(s).

CoNEXT'17,2017,Seoul,Korea

©2017Copyrightheldbytheowner/author(s).

ACMISBN123-4567-24-567/08/06.

https://doi.org/10.1145/nnnnnnn.nnnnnnn

1 TRACEROUTE AND TREENET

tra eroute [10 ℄isone oftheearliestandmostused

topology dis overy tools. It onsists in sending probe

pa kets towards a target with an in reasing

time-to-live (TTL) value until the target replies to the probe.

By olle ting thetime-ex eeded repliesfor ea h TTL

valueand he kingtheirsour eaddress,one an olle t

one IP interfa e of ea h router rossed by thepa kets

sent towards the initial target. Sin e then, variants of

tra eroute have beenproposed to extend its

apabil-ities.Paris tra eroute [1 ℄ an keepa onsistentroute

when tra ing through per-ow load balan ers towards

aspe i targetand anenumerateallpaths[2 ℄.Other

appli ationsoftra eroutein ludessubnetinferen e[9 ℄

aswellasanalyti alalias resolution[5,7℄.

tra eroute also nds appli ation in TreeNET [4 ℄, a

tool whi h maps a target network on the basis of its

subnets. The key idea of TreeNET is to olle t (Paris)

tra eroute measurements towards ea h inferred

sub-netto groupsubnets a ordingto thelast hopin their

respe tiveroute.Doingso,TreeNET anobtaina

subnet-levelmapofthetargetnetwork,whi h analsobeused

to ondu t aliasresolution[3 ℄.Inpra ti e,TreeNET

ob-tainsthesubnetmapbybuildinganetworktree,a

stru -ture where leaves model subnets while internal nodes

model ea h route hop. Ea h internal node is labelled

with an IP interfa e appearing in the routes towards

ea h subnet. With this onstru tion, all subnets

shar-ingthesame lasthop intheirrespe tiverouteends up

underthesameinternalnode.

However,asetof tra eroutemeasurements onsists

mostofthetimein adire ted a y li graph,evenwith

Paristra eroute.Indeed,Paris tra erouteensuresa

route is onsistent for a single target, but tra ing

to-wardsmultipletargetswillstillresultinhavingmultiple

paths. To still beable to build a tree, TreeNET allows

internal nodesto bearmore thanone label(i.e.,an IP

interfa e) [4 ℄. These multi-label nodes therefore model

a olle tionofinterfa esinvolvedinloadbalan ing,and

ensuresea hIPinterfa eobservedwithtra eroute

ap-pearsonlyon eat agivendepthin thetree.

Fig.1(a) showsa simplied topology, while g. 1(b)

(a)Toytopology

(b)Asanetworktree

Figure 1: Prin iples of the network tree of

TreeNET. A,B, C of D are IP interfa esrevealed

by tra eroute.

2 DISCOVERING ALIASES IN

LOAD-BALANCED PATHS

The tree stru ture of TreeNET naturally reveals the

parts of a set of routes that are ae ted by load

bal-an ing.Moreover,TreeNETalreadyimplementsanalias

resolutionmethodologyinvolving IPngerprintingand

several state-of-the-art alias resolution te hniques [3 ℄,

su h as iffinder [6 ℄ and Ally [8℄. Our methodology

therefore simply onsists in applying the alias

resolu-tionmethodologyof TreeNETon thelabels(i.e., IP

in-terfa es)ofmulti-labelnodesappearingin thenetwork

tree.However,werestri ttheobtained aliasesto those

inferredwith iffinderandIP-IDbasedmethods (like

Ally),asthesemethodsareverya urateonsmallsets

ofaddresses. Indeed,other approa hes implemented in

TreeNET might result in overly optimisti results.The

upgradedTreeNETisfreelyavailableonline. 1

3 RESULTS

WedeployedourupgradedTreeNETonPlanetLabto

probe 12 ASes, listed in Table 1, in order to evaluate

therelevan yofaliasresolutiononinterfa esdis overed

in load-balan ed paths. In the next gure, we denote

ea hASbytheindexprovidedinTable1.We olle ted

our dataset on August 27 th

, 2017 and made it freely

available. 2

Themain results of ourmeasurements areshown in

Fig. 2. For ea h AS, we provide three ratios: (

i

proportion of multi-label nodesfor whi h some aliases

wereobtained(white),(

ii

nodesfor whi h only one aliasgathering all labelswas

1

https://github. om/JefGrailet/treenet/v3/

2

Index ASN #IPs Index ASN #IPs

1 37 140,544 7 6453 656,640 2 109 1,173,760 8 8928 827,904 3 703 863,232 9 12956 209,920 4 4711 17,408 10 13789 96,256 5 5400 1,385,216 11 22652 76,288 6 5511 911,872 12 50673 61,696

Table1: TargetASes ofour ampaign

1

2

3

4

5

6

7

8

9

10 11 12

AS index

0

10

20

30

40

50

60

70

80

90

100

Proportion (%)

Nodes w/ aliases

Fully aliased

Aliased/aliasable IPs

Figure2:Statisti sonmulti-labelnodes(August

27 th

,2017).Fromlefttoright:nodeswithaliases,

nodeswherealllabels werealiased togetherand

aliased IP addressesversus aliasableIPs.

inferred (grey),and, (

iii

addresseswithrespe ttothetotalofIPinterfa esthat

ouldbealiased(bla k).Wealsoprovidethes riptfor

generatingsu h agure. 3

The gures shows, via the white bars, that nding

aliasesinmulti-labelnodesisnotsystemati ,evenifwe

have a large amount of aliasable IPs ( f. bla k bars).

However, for almost all ASes, we have one or several

multi-labelnodeswherealllabelswerealiasedtogether.

By looking in depth into the data, wedis overed that

these asesoften orrespondedtomulti-labelnodeswith

fewlabels(lessthan10)thatwerethemselves hildrenof

othermulti-labelnodes.Thissuggestssmallmulti-label

nodes orrespond torouterslo atedat theendof

load-balan edpaths, ausingthemto responddepending on

thepathtaken inthe load balan er.For (mu h)larger

multi-labelnodes,fewaliaseswereobtained.

4 CONCLUSION

In this paper, we took a look at alias resolution on

IPinterfa esfoundin tra eroute measurements

inu-en edbyloadbalan ing.Weproposedasimpleaddition

toTreeNETtodoso,anddeployeditonPlanetLab.Our

rst observations show that, while alias resolution in

su ha ontextdoesnotgivea lotofresults,it ouldbe

usedasawaytospotthe onvergen epointofmultiple

pathsinaloadbalan erandrenesubsequentanalysis.

REFERENCES

[1℄ B.Augustin,X.Cuvellier,B.Orgogozo,F.Viger,T.

Fried-man,M.Latapy,C.Magnien,andR.Teixeira.2006.Avoiding

Tra erouteAnomalieswithParisTra eroute.InPro .ACM

InternetMeasurementConferen e(IMC).

[2℄ B. Augustin, R. Teixeira, and T. Friedman. 2007.

Mea-suring Load-Balan ed Paths in the Internet. In Pro .

ACM/USENIXInternetMeasurementConferen e(IMC).

[3℄ J.-F. Grailet and B. Donnet. 2017. Towards A Renewed

AliasResolutionwithSpa eSear hRedu tionandIP

Finger-printing.InPro .NetworkTra MeasurementandAnalysis

Conferen e(TMA).

[4℄ J.-F. Grailet,F.Tarissan,and B.Donnet.2016. TreeNET:

Dis overingand Conne tingSubnets. InPro . Tra and

MonitoringAnalysisWorkshop(TMA).

[5℄ M. H. Gunes and K. Sara . 2009. Resolving IP Aliases

Transa tionsonNetworking (ToN) 17,6 (De ember2009),

17381751.

[6℄ K. Keys. [n. d.℄. inder. ([n. d.℄). A

tool for mapping interfa es to routers. See

http://www. aida.org/tools/measurement/inder/.

[7℄ K.Keys.2010.Internet-S aleIPAliasResolutionTe hniques.

ACM SIGCOMM Computer Communi ation Review 40, 1

(January2010),5055.

[8℄ N.Spring,R.Mahajan,andD.Wetherall.2002. Measuring

ISPTopologieswithRo ketfuel.InPro .ACMSIGCOMM.

[9℄ M. E. Tozal and K. Sara . 2010. Tra eNET: an Internet

Topology DataColle tor.InPro . ACMInternet

Measure-mentConferen e(IMC).

[10℄ V.Ja obsonetal.1989. tra eroute. manpage.UNIX. See