Measuring IPv6 Performance
Geoff HustonAPNIC Labs
What are we looking at:
• How“reliable”areIPv6connections?
• How“fast”areIPv6connections?
What are we looking at:
• How“reliable”areIPv6connections?
• How“fast”areIPv6connections?
The Measurement Technique
• Embedascriptinanonlinead• HavethescriptgenerateasetofURLstofetch• ExaminethepacketsseenattheservertodeterminereliabilityandRTT
How We Measure
Weuseanonlineadtopresentasequenceofsmallfetchestotheuser’sbrowser
How We Measure
ThesequenceoftestsisusedtotestanumberoftypesofactionsincludingfetchesofIPv4,IPv6andDualstack
How We MeasureWeusetcpdump torecordallpacketactivityattheexperiment’sservers
How Much do We Measure?
What are we looking at:
• How“reliable”areIPv6connections?
• How“fast”areIPv6connections?
Outbound SYN
Busted SYN ACKReturn path
What we see: Connection Failure
serverclient
IPv4 Connection Failure
0.22%FailureRate
IPv4 Failures
• IPv4failuresarearound1in500• AndweareprettysureitsNOT:– Auto-tunnelling– LousyCPEfirmware– Strangefirewallfilters
• Sowhatisthereasonforthisresidualasymmetricfailurerate?• Isitasymmetricroutingconnectivity?
Daily IPv6 Failures
6to4failure:around10%
AverageIPv6failure:around2%
UnicastIPv6failure:around1.5%
Daily IPv6 Failures
Unicast IPv6 shows moderate reliability
1.5%FailureRate
IPv6 Failures
• 1.5%failureforunicastV6isstillunacceptable!• Whyisthishappening?– Auto-tunnelling?– LousyCPEfirmware?– Strangefirewallfilters?– Asymmetricrouting
Is IPv6 failure uniformly distributed?
Is IPv6 failure uniformly distributed?
V6 Failure by Country
V6 Failure by Network
A cautionary note
• Theseare“singleshot”measurementsratherthansustainedrepeatedtest,sothereissomenoisecomponenthere
• Itsalsolikelythatconnectionfailureisrelatedtoconsumerequipmentratherthannetwork-levelfailure,asanetworklevelfailurewouldconventionallygiveafailureratecloserto100%
What are we looking at:
• How“reliable”areIPv6connections?
• How“fast”areIPv6connections?
Let’s look at TCP SYNs
Let’s look at TCP SYNs
1 RTT interval
Why SYNs?
• EveryTCPsessionstartswithaSYNhandshake• Itstypicallyakernelleveloperation,whichmeansthatthereislittleinthewayofapplicationlevelinteractionwiththeSYNexchange
• Onthedownsidethereisonlyasinglesamplepointpermeasurement
Generating a comparative RTT profile
• Foreachsuccessfulconnectioncouplet(IPv4andIPv4)fromthesameendpoint,gatherthepairofRTTmeasurementsfromtheSYN-ACKexchanges• Usetheserver’sweblogstoassociateacoupletofIPv4andIPv6addresses
• UsethepacketdumpstocollectRTTinformationfromtheSYN-ACKExchange
• UseIPv6RTT– IPv4RTTasthemetric
An Example of Path Divergence$traceroutefromSingaporetoCanberra,IPv4tracerouteto202.158.xxxx.yyy,30hopsmax,60bytepackets1103.3.60.30.672ms0.796ms0.899ms2139.162.0.90.754ms0.708ms0.732ms3te0-1-0-21.br03.sin02.pccwbtn.net1.697ms0.760ms0.726ms4ntt.fe3-18.br01.sin02.pccwbtn.net69.526ms69.644ms69.754ms5ae-10.r20.sngpsi05.sg.bb.gin.ntt.net60.702ms68.474ms68.469ms6ae-8.r22.snjsca04.us.bb.gin.ntt.net168.447ms168.532ms168.138ms7ae-19.r01.snjsca04.us.bb.gin.ntt.net167.489ms170.665ms178.832ms8xe-0-8-0-21.r01.snjsca04.us.ce.gin.ntt.net330.084ms323.556ms329.772ms9xe-1-0-1.pe1.msct.nsw.aarnet.net.au330.020ms323.738ms334.474ms10et-3-3-0.pe1.rsby.nsw.aarnet.net.au327.788ms334.157ms328.199ms11138.44.161.6323.644ms319.455ms323.563ms12202.158.xxx.yyy319.885ms333.933ms325.014ms
$traceroute from CanberratoSingapore,IPv4traceroute to139.162.xxx.yyy,64hops max,52byte packets1202.158.x.y0.682ms0.388ms0.313ms2xe-5-0-4-205.pe1.actn.act.aarnet.net.a0.721ms0.828ms0.674ms3et-0-3-0.pe1.rsby.nsw.aarnet.net.au4.548ms4.733ms4.533ms4et-7-1-0.pe1.brwy.nsw.aarnet.net.au4.734ms5.418ms4.745ms5et-0-3-0.pe1.bkvl.nsw.aarnet.net.au5.117ms5.512ms5.524ms6xe-0-0-0.bb1.b.sea.aarnet.net.au148.017ms148.019ms148.131ms7ge3-0.cr02.sea01.pccwbtn.net(206.81.80.13)148.469ms148.059ms148.429ms8tenge0-2-0-14.br03.sin02.pccwbtn.net319.435ms325.053ms319.117ms9tenge0-2-0-15.br03.sin02.pccwbtn.net319.257ms332.560ms323.415ms10linode.te0-1-0-21.br03.sin02.pccwbtn.net323.723ms323.627ms323.587ms11139.162.aaa.bbb334.609ms347.243ms347.220ms12139.162.xxx.yyy325.186ms338.209ms325.603ms
IPv4
We are seeing path symmetry, bouncing off the US West Coast
An Example of Path Divergence
$traceroutefromSingaporetoCanberra,IPv6traceroute6to2001:388:1000:110:e4d:e9ff:x:y,30hopsmax,80bytepackets12400:8901::5287:89ff:fe40:9fc10.897ms0.912ms1.051ms22400:8901:1111::10.851ms0.827ms0.792ms32001:cb0:2102:2:f::10.364ms0.333ms0.516ms42001:cb0:2102:2:f::10.502ms0.461ms0.431ms52001:cb0:21f0:1:17::22.512ms2.176ms3.445ms62001:cb0:21f0:1:17::22.354ms2.382ms1.238ms710gigabitethernet3-5.core1.sin1.he.net1.080ms1.034ms1.020ms810ge1-5.core1.tyo1.he.net88.053ms10ge1-16.core1.hkg1.he.net39.369ms10ge1-5.core1.tyo1.he.net88.084ms
910ge1-5.core1.tyo1.he.net88.157ms100ge8-1.core1.sea1.he.net192.408ms192.642ms
10100ge8-1.core1.sea1.he.net192.631ms192.608ms196.154ms11xe-1-0-1.pe2.brwy.nsw.aarnet.net.au214.176ms186.238ms213.061ms12et-3-1-0.pe1.brwy.nsw.aarnet.net.au211.298ms211.300ms214.200ms13et-1-1-0.pe1.rsby.nsw.aarnet.net.au211.492ms211.359ms211.427ms14et-0-3-0.pe1.actn.act.aarnet.net.au213.332ms211.458ms211.476ms152001:388:1000:110:e4d:e9ff:x.y213.274ms213.199ms213.169ms
$traceroutefromCanberratoSingapore,IPv6traceroute6to2400:8901::f03c:91ff:a:b)64hopsmax,12bytepackets12001:388:1000:110::x:y 0.808ms0.899ms1.586ms2xe-5-0-4-205.pe1.actn.act.aarnet.net.au1.633ms0.646ms0.578ms3et-0-1-0.pe1.dksn.act.aarnet.net.au0.682ms0.649ms0.694ms4et-5-3-0.pe1.crlt.vic.aarnet.net.au8.072ms8.086ms8.049ms5et-5-1-0.pe1.wmlb.vic.aarnet.net.au8.116ms8.055ms8.073ms6et-0-3-0.pe1.adel.sa.aarnet.net.au17.790ms16.984ms17.036ms7et-1-1-0.pe1.prka.sa.aarnet.net.au17.080ms17.152mset-0-3-0.pe1.eper.wa.aarnet.net.au43.319ms
8et-0-3-0.pe1.knsg.wa.aarnet.net.au43.357ms43.443ms43.353ms9gigabitethernet-5-1-0.bb1.b.per.aarnet.net.au43.849ms43.919ms43.850ms10so-0-0-0.bb1.a.sin.aarnet.net.au92.219ms92.275ms92.189ms11as6939.singapore.megaport.com212.347ms212.426ms212.471ms12***132400:8901:1110::2213.924ms213.904ms213.717ms142400:8901::f03c:91ff:a:b213.954ms213.393ms213.726ms
IPv4 – IPv6
In IPv6 we are seeing path asymmetry, with outbound packetsCrossing an Indian Ocean cable, and return packets bouncingoff the US West Coast – the IPv6 RTT is 112 ms smaller than the IPv4 RTT
Global Results
On average IPv6 is showing 20ms – 40ms slower that IPv4
Global Results
Mean Standard Deviation per day of these measurements
Number of sample points per day
RTT Diff Distribution
120 day average
IPv6Faster(ms) IPv4Faster(ms)
RTT Diff by Country
RTT Diff by Network
European RegionV4 faster nets
European RegionV6 faster nets
IPv6 in Spain:
Largest ISPs
IPv6 in Spain:
IPv6 ISPs
IPv6 in Spain – RTT Diff distribution
Germany
German IPv6 ISPs
RTT Diff for DTAG – AS3320
Is IPv6 as “good” as IPv4?
Is IPv6 as “good” as IPv4?
IsIPv6asfast asIPv4?Basically,yesIPv6isfasterabouthalfofthetimeFor75%ofunicastcases,IPv6iswithin10msRTTofIPv4Sotheyperformatmuchthesamerate
(Butthat’sjustforunicastIPv6- theuseof6to4makesthisawholelotworse!)
Is IPv6 as “good” as IPv4?
IsIPv6asrobustasIPv4?IPv4averageconnectionreliabilitycurrentlysitsat0.2%
ThebaseaveragefailurerateofUnicastV6connectionattemptsat1.5%ofthetotalV6unicastconnectionsisnotbrilliant.(6to4isterribleatthis!)
Itcouldbebetter.
Itcouldbeawholelotbetter!
Is IPv6 as “good” as IPv4?
Ifyoucanestablishaconnection,thenIPv4andIPv6appeartohavecomparableRTTmeasurementsacrossmostoftheInternet
ButtheoddsofestablishingthatconnectionarestillweightedinfavourofIPv4!
That’s it!
http://stats.labs.apnic.net/v6perf