perfSONAR Going Forward

perfSONAR GoingForward

EricBoyd,Internet2http://www.perfsonar.net

Internet2TechnologyExchangeSeptember27th 2016

ProblemStatement• TheglobalResearch&Educationnetworkecosystemiscomprisedofhundredsofinternational,national,regionalandlocal-scalenetworks.

September27,2016 2©2016,http://www.perfsonar.net

ProblemStatement• Whilethesenetworksallinterconnect,eachnetworkisowned

andoperatedbyseparateorganizations(called“domains”)withdifferentpolicies,customers,fundingmodels,hardware,bandwidthandconfigurations.

September27,2016 3©2016,http://www.perfsonar.net

ProblemStatement• Thiscomplex,heterogeneoussetofnetworksmustoperateseamlesslyfrom“endtoend”tosupportscienceandresearchcollaborationsthataredistributedglobally.

September27,2016 4©2016,http://www.perfsonar.net

ProblemStatement• Inpractice,performanceissuesareprevalentanddistributed.• Whenanetworkisunderperformingorerrorsoccur,itis

difficulttoidentifythesource,asproblemscanhappenanywhere,inanydomain.

September27,2016 5

• Local-areanetworktestingisnotsufficient,aserrorscanoccurbetweennetworks.

©2016,http://www.perfsonar.net

ProblemStatement:Hardvs.SoftFailures

• “Hardfailures”arethekindofproblemseveryorganizationunderstands– Fibercut– Powerfailuretakesdownrouters– Hardwareceasestofunction

• Classicmonitoringsystemsaregoodatalertinghardfailures– i.e.,NOCseessomethingturnredontheirscreen– Engineerspagedbymonitoringsystems

September27,2016 6©2016,http://www.perfsonar.net

ProblemStatement:Hardvs.SoftFailures

• “Softfailures”aredifferentandoftengoundetected– Basicconnectivity(ping,traceroute,webpages,email)works

– Performanceisjustpoor

• Howmuchshouldwecareaboutsoftfailures?

September27,2016 7©2016,http://www.perfsonar.net

September27,2016 8©2016,http://www.perfsonar.net

Physicalpipethatleakswateratrateof.0046%byvolume.

è è

Network‘pipe’thatdropspacketsatrateof.0046%.

è è

Result100%ofdatatransferred,slowly,at<<5%optimalspeed.

Elephant Flows Place Great Demands on NetworksResult99.9954%ofwatertransferred,at“linerate.”

essentiallyfixed

determinedbyspeedoflight

Throughcarefulengineering,wecanminimizepacketloss.

September27,2016 9©2016,http://www.perfsonar.net

SoftFailuresCausePacketLossandDegradedTCPPerformance

MetroArea

Local(LAN)

RegionalContinental

International

Measured (TCP Reno) Measured (HTCP) Theoretical (TCP Reno) Measured (no loss)

With loss, high performance beyond metro distances is essentially impossible

September27,2016 10©2016,http://www.perfsonar.net

publicperfSONARServers(Sept2016)• Over2000publiclyregisteredservers

– Equalnumberofnon-registeredservers?• ESnet:50

– mostly10G,includesa40GhostinBoston• GEANT:22• Internet2:3• Someothertopdeployments:

– Onenet (24),AMPATH(8),bc.net (10),RNP(8),Canarie (13),kreonet(14),NERO(12),AARnet (19),JGN(17),CENIC(5),KANREN(5)

September27,2016©2016,http://www.perfsonar.net 11

MoreperfSONARStatistics• TotalHosts:2001• TotalDomains:413• 75%arerunninglatestversion(auto-update)• 7%arerunningaversion<3.5(end-of-life,orphanhosts?)

• 27%ofthehostshaveanIPV6Address• 38%are.edu hosts• 75totaltop-leveldomains,940domains

September27,2016©2016,http://www.perfsonar.net 12

MoreperfSONARStats• 95%RHEL/CentOS;5%Debian/Ubuntu• 10%areVMs• 40%usingjumboframes• NICSpeed:

– 1Gbps49.06%– 10Gbps44.17%– 40Gbps2.88%– 100Mbps 2.32%

September27,2016©2016,http://www.perfsonar.net 13

perfSONAR 4.0• perfSONAR 4.0FeatureTour

– AndyLake– 9:00AM

• pScheduler DeepDive– MarkFeit– 9:30AM

September27,2016©2016,http://www.perfsonar.net 14

Whatcomesnext?• perfSONAR SteeringGroupisengagedinaprojecttodefinestrategicplanforpost-perfSONAR 4.0efforts

• Fourbasicthemesarebeingconsidered:– Operationsefficiency– Automationofconfiguration,execution,&analysis– Performanceofthecloud– Careandfeedingoftheopensourceeffort

September27,2016©2016,http://www.perfsonar.net 15

OperationsEfficiency• Reduceoperationseffortrequiredofnetworkoperatorsanddataintensivesciencecollaborations– EasydeploymentofephemeralpS Nodes:perfSONARcontainerization(e.g.Docker)

– Automateloganalysistoidentifykeysignatures– Improvedashboardvisualization– Enhancemeshconfigurations

AutomationofConfiguration,Execution,&Analysis

• Configuration– Automatictestconfiguration

• SignificantprogressinperfSONAR 4.0– Automaticnodediscovery/insertionintotestmessages

• Execution– Auto-detectionofproblemsalongtheend-to-endpath– Automaticallyrunadditionaltests

• Analysis– Anomalydetectiontools– Intelligentalarming(e.g.adaptiveviamachinelearning)– Archivesforresearchdata

PerformanceoftheCloud• Cloudservicesandcloudbasedworkflowsareimportantgoingforward.

– Toourresearchers– Toourlabsanduniversities

• Performancemonitoringneedstobeautomatedand‘ephemeral’tillthecloudworkflowisactive

• Supportatargetedpopularusecase– e.g.:perfSONAR AmazonEC2deployment– Showcasewhatispossible– Researchhowtosolvetheproblemmoregenerally

• Buildtowardsamoregeneralsolutioninfollow-onreleases• Observation:Increasingneed/useofperfSONAR onVMs

– Informtheuserofthelimitationsofwhatcanbeconcluded

CareandFeeding• Expandnewofferings

– pScheduler isanewoffering(1.0release)whichwillpromptanewiterationofcommunityideas• Securitychallenges,vulnerabilities,andinefficiencies

– e.g.Expandunittesting– Review,refine,andreduceinefficientcodecomponents

• Supportevolvingdefaultoperatingsystems– e.g.CentOS6->Centos7,Debian 7->Debian 8->Debian 9

• PreferredAPIinterfaces,middlewarekeepevolving– e.g.SOAP/XML->REST/JSONAPIs– messagepassingarchitectures– e.g.RabbitMQ

• CustomCodeReplacement– RemoveperfSONAR customcodewithwell-supportedpackagesastheybecomeavailable

• ELK(Elasticsearch,Logstash,Kibana)• ESnet’s ReactTimeseries Charts• IU’sTimeSeriesDataService(TSDS)

• SmallNodes– Emergingtechnologiesarerapidlyevolving– Maintainacurrentsnapshotofbestofferings

CommunityMilestones&Feedback• Milestones

– perfSONAR 4.0– RC1isoutthisweek– perfSONAR 4.0– RC2expectedinOctober– perfSONAR 4.0outinNovember

• Feedback– WhatdoyouneedfromperfSONAR post4.0?

September27,2016©2016,http://www.perfsonar.net 20