p4 deep dive - netlab€¦ · p4 three goals 1. protocol independence • no native support for any...

[email protected]

7th September2017

Agenda

• 9:30– P4motivation,languageoverview• OpenFlow vsP4• SDNtoday

• ~10:30– P4demos,exampleapplications• Ipv4forwardingswitch• Calculator(run1+1onyourrouter!)• Multi-hoprouteinspection(in-networktelemetry)• Trackingper-flowTCPretransmissions

• Discussions:dataplaneprogrammability@ESnet

04/10/2017 P4deepdive@ESnet- RichardCziva 2

TheroadtoP4P4deepdive


Software-DefinedNetworking

• Programmablecontrolovernetworks• Physicallyseparatedcontrolplanetheforwardingplane• MostpopularrealizationofSDN’sdata-planeabstractionisOpenFlow• OpenFlow wasborn10yearsago

• Andalothaschanged


Software-DefinedNetworking

• OpenFlow becamethedefactostandardfornetworkprogramming• OriginallyitwastargetingLANandDCnetworks• LatelytheykeptextendingOpenFlow tosupportnewprotocols(mostlyencapsulation,networkvirtualization)• VxLAN (2010)• STT(StatelessTransportTunnel)(2012)• NSH(2014)• …


OpenFlow matchfields


• Matchfieldskeepgrowingandgrowing:hardforvendorstokeepup

Networkoperatorscanhavespecificrequirements!Canweletthemimplementit?

Apartialsolution:BPFmatching

• InsteadoffixedOpenFlow matchfields,whydon’tweuseaBerkeleyPacketFilters (BPF)?• (TheoriginalBPFpaperwaswrittenby StevenMcCanne and VanJacobson in1992whileat LawrenceBerkeleyLaboratory[9][10])

• BPFcanbeusedtodescribeapacketmatchingasandirectedacyclicgraph


Apartialsolution:BPFmatching

• ABPFprogramwassmallenoughtobeplacestoanOpenFlow packet(OXMfield)• BPFcanbeexecutedonanyplatform(platform-independence)• BPFreliesonsimpleinstructions(load,compare)->easytoimplementonhardware


Networkdata-planetoday

• TheASICdefineshowcanwehandlepackets• ASICsareallclosed,wecan’textendthemifrequired• Networksaredesignedusingabottom-updesign


FixedfunctionASIC

SwitchOS

API

driver

Networkingtomorrow

• Wewanttotellthedevicehowtobehave(toptobottomapproach)• Programmablenetworkdevices:• PISA(FlexibleMatch+Action ASICs):IntelFlexpipe,BarefootTofino• NPU:Netronome• CPU:DPDK,eBPF,OpenvSwitch• FPGA:XilinxNetFPGA

• P4definesalanguagetoprogramsomeofthesedevices


Programmableswitch

SwitchOS

API

driver

Benefitsoftopàbottom approach

• Tonetworkproviders• Eliminating”blackboxes”and“magicbehavior”• Developerscanprogram,testanddebugnetworkdevicesallthewaydown• Customrouting/switchingprotocolscanbekept“in-house”

• Tovendors• Extremelyfastiterationandfeaturerelease• Newbusinessmodelsforsellingsoftwareandhardwareseparate• Vendorscanfixdata-planebugsafterdeviceshavebeendeployed• Softwaredevelopmentpracticesusedineveryphase


Agenerallanguage:P4

• P4:alanguagedesignedtoallowtoptobottomprogrammability• P4wasproposedatSIGCOMM2013:

Bosshart,Pat,DanDaly,GlenGibb,MartinIzzard,NickMcKeown,JenniferRexford,ColeSchlesingeretal."P4:Programmingprotocol-independentpacketprocessors." ACMSIGCOMMComputerCommunicationReview 44,no.3(2014):87-95.

• P4consortium:over80vendors,manyUniversities,individuals…


P4vsOpenFlow


P4inveryhigh-levelP4deepdive


P4threegoals

1. Protocolindependence• Nonativesupportforanyprotocol• Theprogrammerdescribestheprotocols

2. Targetindependence• P4programsaredesignedtobe implementation-independent• CanbecompiledtoCPUs,NPUs,FPGAs,ASICs,networkprocessors,etc

3. In-fieldre-configurability• Duetotargetandprotocolindependenceandtheabstractlanguage,P4targets(e.g.,ASICs)canchangetheirbehaviorovertheirlifetime


P4vstraditionalswitches

• InaP4definedswitchthedataplaneisnotfixed,butdefinedbyaP4program• InaP4definedswitchtheAPIbetweenthedataplaneandthecontrolplaneisdefinedbyaP4program


From:https://p4lang.github.io/p4-spec/docs/P4-16-v1.0.0-spec.html

P4pipeline

• Parser:Parsesrawpacketsfromwireintometadata• Match+Actiontables:operateonmetadata• Deparser:Serializesmetadataintoapacket• Metadatabus:availableintheentirepipeline


Table1 Table2 Tablen

Deparser

Parser

Metadatabus

packet packet

AnexamplepipelinefromCorsa


CouldbeexpressedwithP4!

Languageoverview(P4_16)P4deepdive


Languageversions

• Therearemultipleversionsofthelanguage• Unfortunatelytheversionsarebackwards-incompatible• Andsomethingsonlyworkinformerversionsjustnow

• Iamintroducingthelatest:P4_16• Smallercorelanguagethanpreviousversions(only40keywords)• Thisissupposedtobeastablelanguagedefinition



Coreabstractions

• Headertypes:definehowyourprotocolheaderslooklike• Parsers:tellthemhowtoparseheaders• Tables:containstateassociatinguser-definedkeyswithactions• Actions:describinghowapacketshouldbemanipulated• Match-Actionunits:createlookupkey,performtablelookup,executeaction• Controlflow:specifiestheorderofapplyingmatch-actionunits• Externobjects:registers,counters,meters,etc• Metadata:datastructuresassociatedwitheachpacket


Headerdefinitions

• Allowsyou,thedevelopertodescribepacketformatsandnameheaderfields• Fixedvsvariablelengthfieldsarebothpermitted


header Ethernet_h {bit<48>dstAddr;bit<48>srcAddr;bit<16>etherType;

}

header IPv4_h{bit<4>version;bit<4>ihl;bit<8>diffserv;bit<16>totalLen;bit<16>identification;bit<3>flags;bit<13>fragOffset;bit<8>ttl;bit<8>protocol;bit<16>hdrChecksum;bit<32>srcAddr;bit<32>dstAddr;varbit<320>options;

}

Parser

• Afinitestatemachinewithonestartstateandtwofinalstates


Start

Ethernet?

TCP UDP

Accept

Reject

IfEthernet

IfTCP

else

else

IfUDP

parserMyParser(){statestart{transitionparse_ethernet;}stateparse_ethernet {

packet.extract(hdr.ethernet);transitionselect(hdr.ethernet.etherType){

TYPE_IPV4:parse_ipv4;default:accept;

}}…}

Tables

• Tablescontainstate(pushedbythecontrolplane)toforwardpackets• Tablesdescribeamatch-actionunit• Packetmatchingcanbedoneeither:• Exactmatching• LongestPrefixMatch(LPM)• Ternarymatching(masking)

• Allpossibleactionshavetobedefinedinadvance


tableipv4_lpm{reads {

ipv4.dstAddr:lpm;}actions {

forward()}

}

Match-Actionunits



Actions

• Actionsconsistofcodeandtakeactiondata• Actiondatacomesfromthecontrolplane(e.g.,IPaddresses/portnumbers)

• Specific,loop-freeprimitivescanbeexecutedinanaction• Numberofinstructionsmustbepredictable,sonoloopsorconditionalstatementsallowedhere


action ipv4_forward(macAddr_t dstAddr,egressSpec_t port){standard_metadata.egress_spec =port;hdr.ethernet.srcAddr =hdr.ethernet.dstAddr;hdr.ethernet.dstAddr =dstAddr;hdr.ipv4.ttl=hdr.ipv4.ttl- 1;

}

Externs

• ExternsarearchitecturespecificconstructswithwelldefinedAPIs• Examples:• Checksumcalculation• Registers• Counters• Meters


extern Checksum16{Checksum16();//constructorvoid clear();//prepareunitforcomputationvoid update<T>(in Tdata);//adddatatochecksumvoid remove<T>(in Tdata);//removedatafromexistingchecksumbit<16>get();//getthechecksumforthedataaddedsincelastclear

}

extern register<T>{register(bit<32>size);void read(outTresult,inbit<32>index);void write(inbit<32>index,inTvalue);

}

Metadata

1. User-definedmetadata(emptystruct atfirstforeachpacket)• Youcanputhereanythingyouwant• Accessibleeverywhereinthepipeline• Usefule.g.,forstoringthepackethash

2. Intrinsicmetadata- thisisprovidedbythearchitecture• Ingressport,egressportisdefinedhere• Timestampwhenpacketwasqueued,queuedepth• Multicasthash/multicastqueue• Packetpriority,packetcolor• Egressportspecification(e.g.,outputqueue)


Controlflow

• Stitchingeverythingtogether• Imperativeprogramexpressingthehigh-levellogicandthesequenceformatch-actionunitinvocations


control ingress(){[tabledefinitionshere][actiondefinitionshere]apply { if (hdr.ipv4.isValid()){ipv4_lpm.apply();

} }

}

V1Switch(ParserImpl(),verifyChecksum(),ingress(),egress(),computeChecksum(),DeparserImpl())main;

P4compiler,toolsP4deepdive


P4compiler

• TheP4compiler(p4c)produces:1. Dataplaneruntime2. APIformanagingthestateinthedataplane


P4softwareswitch

• p4lang/p4c-bm:generatesJSONconfigurationforbmv2• p4lang/bmv2:softwareswitchthatunderstandstheJSONconfig• bmv2:behaviormodelversion2


myprogram.p4 myprogram.json

bmv2-ssload

Compilewithp4c-bm

Dataplane

API

Runtimecommands

RuntimeAPIforsoftwareswitch

• Allowsyoutomanipulatetables,readregisters,counters…

• APIiseasilyusablewiththeprovidedsimple_switch_CLI program


- table_set_default <tablename><actionname><actionparameters>- table_add <tablename><actionname><matchfields>=><actionparameters>[priority]- table_delete <tablename><entryhandle>

p4@p4d2:~$simple_switch_CLI --thrift-port9090ObtainingJSONfromswitch...DoneControlutilityforruntimeP4tablemanipulationRuntimeCmd:counter_read counter0thisisthedirectcounterfortableipv4_lpmcounter[0]=BmCounterValue(packets=4,bytes=88)

Switchlogs+debugger

• Simpleswitchlogsareusefulenough:

• Thereisalsoagdb likedebugger:• Breakpoint,watchingaheaderfield,printsthevalueofaheaderfield,etc


p4@p4d2:~$tail-f/home/p4/p4_tutorial/P4D2_2017/exercises/mri/build/logs/s1.log[20:58:45.689][bmv2][D][thread1572][94.0][cxt 0]Pipeline'ingress':start[20:58:45.689][bmv2][D][thread1572][94.0][cxt 0]Pipeline'ingress':end[20:58:45.689][bmv2][D][thread1572][94.0][cxt 0]Egressportis0….[22:28:07.263][bmv2][T][thread1572][166.0][cxt 0]Applyingtable'ipv4_lpm’[22:28:07.263][bmv2][D][thread1572][166.0][cxt 0]Lookingupkey:

*ipv4.dstAddr:0a00020a[22:28:07.263][bmv2][D][thread1572][166.0][cxt 0]Table'ipv4_lpm':hitwithhandle1

P4examplesP4deepdive


Calculator

• Whosaidanetworkdevicecannotbeusedasacalculator?• Idea:• wesendpacketsdescribinganoperation(e.g.,1+1)• thedeviceperformstheoperationandwritestheresulttothepacket• thedevicesendsthepacketbacktothesenderthatparsestheresult(2)


H1 P4switch1+1?

2

Asimpleswitch

• BasicIPv4forwardingswitch• Stepbystep:

1. Packetisparsed(Ethernet,IPv4only)2. PacketismatchedusingLPMonIPv4

destinationaddresses3. Egressportisset4. Packet’sEthernetsourceanddestination

isset5. TTL=TTL-16. Checksumrecalculation

• Additional:counters


h1

h2

h3

s1

s2

s3

Multi-HopRouteInspection(MRI)

• Let’sfindoutwherehasourbeeninthenetwork• WewilluseTCP’soptionsfieldforthedata(parsingastackofheaders)• Everyswitchwillwriteit’sownID(swid)tothepacket• PacketfromH1-->H2willcollectswids[1,2]


h1

h2

h3

s1

s2

s3

TCPretransmissiontracker

• Goal:IdentifyallflowsthatareexperiencingTCPretransmission1. TrackifwehaveseenthesameTCP

packetbefore(SYNnumber)1. BasedonRTO(min1s),exponentialbackoff

2. TrackSACKflags• BetterwithburstsofpacketsinoneRTO

• ”Controller”:readsthecountersfromthedata-plane


Client Server

Seq =1

Seq =1Ack =1

RTO

Client Server

Seq =2

Seq =1Ack =1

RTOSeq =3

Ack =1,Sack=3Seq =2

PISA/PSA

• PSA=PortableSwitchArchitecture– WGinP4consortium• PISA =ProtocolIndependentSwitchArchitecture– referencechipdesign• ReferencearchitectureforaP4programmableswitch• Configurableparser,tables,etc.• BarefootTofinoisafullyprogrammablePISA• “World’sfastestandmostprogrammableswitchseriesupto6.5Tbps”

• Actualdevice:Edgecore Wedge100BF-65X:2RU65x100GE,6.5Tbit/s


Takeaway

• P4advocatesatop-downapproachfornetworkprogramming• Insteadofleavingthedevicestellushowtheyprocesspackets=>wewouldliketotellthemhowtodoit

• P4isjustonestandard,manymoretocome,however:


Inthenextfewyearsdata-planeprogrammabilitywillbecomecommonplaceandP4playsanimportantroleinit

References

• Figuresandsomedemoshavebeenselected:• P4tutorialslidesatSIGCOMM:https://github.com/p4lang/tutorials/blob/master/SIGCOMM_2016/p4-tutorial-slides.pdf• P4languagedocs:https://p4lang.github.io/p4-spec/docs/P4-16-v1.0.0-spec.html• ExamplesfromGithub:https://github.com/p4lang/tutorials

• P4paper


Thankyou

Thankyouforyourattention!


p4 deep dive - netlab€¦ · p4 three goals 1. protocol independence • no native support for any...

Documents