PISCES:'A'Programmable,'Protocol4 Independent'So8ware'Switch' [SIGCOMM'2016]'

Transcription

1 PISCES:AProgrammable,Protocol4 IndependentSo8wareSwitch [SIGCOMM2016] Sean%Choi% SlideCreditsto ProfessorNickMcKeownandMuhammadShahbaz

2 Outline MoLvaLonsandhistoryofSDN UsecasesofSDN SDNandthechangeinthenetworkingstack WhatisP4and ProtocolIndependentPacketProcessing? IntroducingPISCES 2

3 Outline MoLvaLonsandhistoryofSDN UsecasesofSDN SDNandthechangeinthenetworkingstack WhatisP4and ProtocolIndependentPacketProcessing? IntroducingPISCES 3

4 What is Software-Defined Networking (SDN)? 4

5 Software Defined Network Anetworkinwhichthecontrolplaneis physicallyseparatefromtheforwardingplane. and$ Asinglecontrolplanecontrols severalforwardingdevices. (That sit) 5

6 So8wareDefinedNetwork(SDN) Control Program Control Program Global Network Map ControlPlane Control Program Control Packet Forwarding Control Packet Forwarding Control Packet Forwarding Control Packet Forwarding Control Packet Forwarding 6

7 So8wareControlPlane SDN Control Packet Forwarding Control Packet Forwarding Control Packet Forwarding Control Packet Forwarding Control Packet Forwarding Intendedconsequences Putnetworkownersandoperatorsincontrol. 2. Networksthatcostless:simpler,streamlinedhardware. 3. Networksthatcostlesstooperate(fewerfeatures). 4. Networksthatevolvefaster. 7

8 OriginsofSDN MarLnCasado TheEthaneProject [SIGCOMM2007] 8

9 2006 Howdifficultisittodefineallnetwork operalonsinso8ware,outsidethedatapath? Stanfordcampus 35,000 users 10,000 new flows/sec 137 network policies 2,000 switches 2,000 switch CPUs 9

10 CrazyquesLon:Whatifso8waredecideswhether toaccepteachflow,andhowtorouteit? Controllers 1. Checkflowagainstpolicy 2. Pickroute A B 10

11 Howmany$400controller serversdoweneedtoservice 35,000users? 11

12 LessthanOne 12

13 Ifwecandefinenetworkbehavioroutsidethe datapath,theneventuallywewill. 13

14 Whathappenednext SDN,OpenFlow,OpenvSwitch,NetworkVirtualizaLon, About250startupssofar. Source:SDXCentral 14

15 Outline MoLvaLonsandhistoryofSDN UsecasesofSDN SDNandthechangeinthenetworkingstack WhatisP4and ProtocolIndependentPacketProcessing? IntroducingPISCES 15

16 SDN use cases Routing and NFV 16

17 function Dijkstra(Graph, source): for each vertex v in Graph: dist[v] := infinity ; previous[v] := undefined; dist[source] := 0 ; Q := the set of all nodes in Graph ; while Q is not empty: // The main loop u := vertex in Q with smallest distance in dist[] ; remove u from Q ; if dist[u] = infinity: break ; Edsger Dikjstra for each neighbor v of u: alt := dist[u] + dist_between(u, v) ; if alt < dist[v]: dist[v] := alt ; previous[v] := u ; decrease-key v in Q; return dist[], previous[]; end function 17

18 18

19 1. Figure out which routers and links are present. 2. Run Dijkstra s algorithm to find shortest paths. If a packet is going to B, then send it to output 3 Data B B 19

20 95% 5% 1. Figure out which routers and links are present. 2. Run Dijkstra s algorithm to find shortest paths. 50,000 lines of code B 20

21 % Dijkstra 5%% Network 95%% Map Dijkstra Global Network Map NetworkOS OS Specialized Hardware Packet Forwarding Packet Forwarding Packet Forwarding Packet Forwarding 21

22 NetworkFuncLonVirtualizaLon(NFV) Middlebox Middlebox Middlebox Middlebox PublicInternet Packet Forwarding Packet Forwarding Packet Forwarding Packet Forwarding Packet Forwarding 22

23 NetworkFuncLonVirtualizaLon(NFV) VM VM VM VM VM VM Packet Forwarding Middlebox PublicInternet Packet Forwarding Packet Forwarding Packet Forwarding Packet Forwarding Packet Forwarding 23

24 Dijkstra IS4IS BGP MPLS NFV Global Network Map NetworkOS 24

25 Outline MoLvaLonsandhistoryofSDN UsecasesofSDN SDNandthechangeinthenetworkingstack WhatisP4and ProtocolIndependentPacketProcessing? IntroducingPISCES 25

26 Specialized Features Specialized Control Plane Specialized Hardware Hundredsofprotocols 7,000RFCs Tensofmillionsoflinesofcode. Closed,proprietary,outdated. Billionsofgates. Powerhungryandbloated. 26

27 Specialized Applications Specialized Operating System Specialized Hardware App App App App App App App App App App App Open Interface Windows Linux (OS) or or Open Interface Microprocessor Mac OS Vertically integrated Closed, proprietary Slow innovation Horizontal Open interfaces Rapid innovation 27

28 Specialized Applications Specialized Operating System AppAppAppAppAppAppAppAppAppAppApp Open Interface Control Control Plane or Plane or Control Plane Open Interface Specialized Hardware Vertically integrated Closed, proprietary Slow innovation Merchant Switching Chips Horizontal Open interfaces Rapid innovation28

29 Icancustomizemynetworks!...? 1. Seewhatmyforwardingplaneisdoing. 2. Quicklydeploynewprotocols 3. PutexpensivemiddleboxfuncLonsintothenetwork. 4. DifferenLate.NowIownmyintellectualproperty. 5. TryoutbeauLfulnewideas.Tailormynetworktomeetmyneeds. NotReally 29

30 WhataboutthefixedfuncLonswitch? SwitchOS Run4LmeAPI Driver ThisishowIprocesspackets FixedfuncLonswitch 30

31 Problems:FixedfuncLonswitch 1. SlowinnovaLon Severalmonthsoryearstoaddanewfeatureorprotocol 2. Inefficient Matchtableshard4wiredtospecificpurpose 3. Complicated Switchimplementssupersetofallfeatures 4. Leadstoborom4updesign FrustraLngforprogrammers 31

32 Whatwewant:ProgrammableSwitch Thisishowtheswitchmust processpackets SwitchOS Run4LmeAPI Driver P4canhelpusdothis! ProgrammableSwitch 32

33 Outline MoLvaLonsandhistoryofSDN UsecasesofSDN SDNandthechangeinthenetworkingstack WhatisP4and ProtocolIndependentPacketProcessing? IntroducingPISCES 33

34 P4 P4: ProgrammingProtocol4Independent PacketProcessors ACMCCR.Volume44,Issue#3(July2014) PatBosshart,GlenGibb,MarLnIzzard,andDanTalayco(BarefootNetworks), DanDaly(Intel),NickMcKeown(Stanford),ColeSchlesingerandDavidWalker (Princeton),AminVahdat(Google),andGeorgeVarghese(Microso8) 34

35 PhasesforProtocol4Independent PacketProcessing Phase%0.IniLally,theswitchdoesnotknowwhata protocolis,orhowtoprocesspackets (ProtocolIndependence) Phase%1.Wetelltheswitchhowwewantittoprocess packets(configuralon) Phase%2.Theswitchruns(Run4Lme) 35

36 ThreeGoals Protocolindependence Configureapacketparser Defineasetoftypedmatch+acLontables Targetindependence Programwithoutknowledgeofswitchdetails Relyoncompilertoconfigurethetargetswitch Reconfigurability Changeparsingandprocessinginthefield 36

37 37

38 The Abstract Forwarding Model IniLally,aswitchisunprogrammed anddoesnotknowanyprotocols. Parser Match+AcLonTables Queues/ Scheduling PacketMetadata 38

39 1 Protocol Authoring L2_L3.p4 SwitchOS 4 Run! 2 Compile 3 Configure Eth IPv4 VLAN IPv6 Run4LmeAPI Driver TCP Parser New Match+AcLonTables Queues/ Scheduling PacketMetadata 39

40 1 Protocol Authoring L2_L3.p4 OF143.p4 SwitchOS 4 Run! 2 Compile 3 Configure Run4LmeAPI Driver Parser Match+AcLonTables Queues/ Scheduling PacketMetadata 40

41 HeadersandFields TheParser Match+AcLonTables Controlflow P4inDetail 41

42 HeadersandFields HeaderFields:Ethernet header_type ethernet_t { fields { dstaddr : 48; srcaddr : 48; ethertype : 16; } } /* Instance of eth header */ header ethernet_t first_ethernet; Metadata header_type standard_metadata_t { fields { ingress_port : 32; packet_length : 32; ingress_timestamp : 32; egress_spec : 32; egress_port : 32; egress_instance : 32; } } metadata standard_metadata_t std_metadata; 42

43 TheParser Parser:Ethernet parser parse_ethernet { extract(ethernet); return switch(latest.ethertype) { ETHERTYPE_VLAN : parse_vlan; ETHERTYPE_MPLS : parse_mpls; ETHERTYPE_IPV4 : parse_ipv4; ETHERTYPE_IPV6 : parse_ipv6; ETHERTYPE_ARP : parse_arp_rarp; ETHERTYPE_RARP : parse_arp_rarp; } } Parser:IPv4 parser parse_ipv4 { extract(ethernet); return switch(latest.ethertype) { PROTO_TCP : parse_tcp; PROTO_UDP : parse_udp;... } } 43

44 Match+Action Tables Specifies Whichfieldstoexamineineachpacket AcLonsthatmaybeapplied(byrule) Tablesize(opLonal) Match+AcLonTable:VLAN table port_vlan { reads { std_metadata.ingress_port : exact; vlan_tag[outer_vlan].vid : exact; } actions { drop, ing_lif_extract; } size 16384; } Match+AcLonTable:UnicastRPF table urpf_check { reads { routing_metadata.bd : ternary; ipv4.dstaddr : ternary; } actions { urpf_clear, urpf_set; } } 44

45 AcLons BuiltfromprimiLves modifyfield(packetheaderormetadata) add/removeheader clone/recirculate counter/meter/statefulmemoryoperalons Parallel semantics AcLons:LIFExtract /* Ingress logical interface setup */ action ingress_lif_extract(i_lif, bd, vrf, v4term, v6term, igmp_snoop) { modify_field(route_md.i_lif, i_lif); modify_field(route_md.bd, bd); modify_field(route_md.vrf, vrf); modify_field(route_md.ipv4_term, v4term, 0x1); modify_field(route_md.ipv6_term, v6term, 0x1); modify_field(route_md.igmp_snoop, igmp_snoop, 0x1); } 45

46 ControlFlow ControlFlow:Ingress control ingress { apply_table(port); apply_table(bcast_storm); apply_table(ip_sourceguard); if (valid(vlan_tag[0])) { apply_table(port_vlan); } apply_table(bridge_domain); if (valid(mpls_bos)) { apply_table(mpls_label); } retrieve_tunnel_vni(); if (valid(vxlan) or valid(genv) or valid(nvgre)) { apply_table(dest_vtep); apply_table(src_vtep); }.... } 46

47 TheP4ViewoftheWorld P4%Program/Library% Compiler Compiler Compiler Acompilerpertarget SwitchconfiguraLon Switch Switch Switch PacketForwarding PacketForwarding PacketForwarding Engine Engine Engine 47

48 PISCES:FirstEverP4tovSwitchCompiler P4%Program/Library% PISCESCompiler OpenvSwitch PacketForwarding Engine 48

49 Outline MoLvaLonsandhistoryofSDN UsecasesofSDN SDNandthechangeinthenetworkingstack WhatisP4and ProtocolIndependentPacketProcessing? IntroducingPISCES 49

50 ImportanceofSo8wareSwitches VM* VM* VM* VM* VM* VM* VM* VM* VM* VM* VM* VM* VM* VM* VM* VM* OVS$ Hypervisor* OVS$ Hypervisor* OVS$ Hypervisor* OVS$ Hypervisor* ToR* ToR* ToR ToR Core* 50

51 ImportanceofSo8wareSwitches VM* VM* VM* VM* VM* VM* VM* VM* VM* VM* VM* VM* VM* VM* VM* VM* OVS$ Hypervisor* OVS$ Hypervisor* OVS$ Hypervisor* OVS$ Hypervisor* ToR* ToR* ToR ToR Core* 51

52 EaseofCustomizaLon? VM* VM* VM* VM* OVS$ Hypervisor* EnableRapid%Developmentand DeploymentofNetwork%Features!% Core* ToR* Is%it%REALLY%the%case?% 52

53 EaseofCustomizaLon? VM* VM* VM* VM* OVS$ Hypervisor* ToR* Core* VM* VM* VM* VM* OVS$ Hypervisor* ToR* Forexample,OVSsupportsfollowing tunnelingprotocols: 4 VXLAN:VirtualExtensibleLAN 4 STT:StatelessTransportTunneling 4 NVGRE:NetworkVirtualizaLon GenericRouLng 53

54 RapidDevelopment&Deployment? OVS$ Kernel* Fast*Packet*IO*(or* DPDK* Forwarding)* 54

55 RapidDevelopment&Deployment? OVS$ Parser* Packet*Processing*Logic* MatchCAction*Pipeline* Kernel* DPDK* Requires%domain%experTze%in:% 4 Networkprotocol%design 4 So8waredevelopment 4 Develop% 4 Test% 4 Deploy% large,complexcodebases. Arcane%APIs% 4 Cantake3Y6%monthstogetanew featurein. 4 Maintaining%changesacrossreleases 55

56 RapidDevelopment&Deployment? OVS$ Parser* MatchCAction*Pipeline* Kernel* DPDK* 56

57 RapidDevelopment&Deployment? OVS$ Parser* MatchCAction*Pipeline* Kernel* DPDK* 57

58 RapidDevelopment&Deployment? P4$ Parser* MatchCAction*Pipeline* 341linesofcode OVS$ Compile* NaTve%OVS Parser* MatchCAction*Pipeline* Kernel* DPDK* 14,535%linesofcode 58

59 RapidDevelopment&Deployment? P4$ Parser* MatchCAction*Pipeline* Compile* Performance%overhead!% OVS$ Parser* MatchCAction*Pipeline* Kernel* DPDK* 59

60 What sthecostofprogrammabilityon Performance? 60

61 PISCES:AProtocol4IndependentSo8wareSwitch P4* PISCES$ vswitch* OVS* 61

62 PISCES:AProtocol4IndependentSo8wareSwitch P4* Runtime*Flow*Rules* $ Compiler* * Flow*Rule*Checker* parse* match* action* OVS* Executable* 62

63 PISCES:AProtocol4IndependentSo8wareSwitch 4 P4andOVSpacket%forwarding%models. 4 Performance%overheadofanaïve%mappingfrom P4toOVS. 4 PISCES%compiler%opTmizaTonstoreducethe performanceoverhead.

64 P4ForwardingModel(orPost4PipelineEdiLng) Ingress* Packet* Parser* MatchC Action* Tables* Packet* Deparser* Egress* Ingress*Packet* Egress*Packet* Header* Fields* 64

65 OVSForwardingModel(orInlineEdiLng) Ingress* Packet* Parser* MatchC Action* Tables* Egress* Ingress*Packet* Egress*Packet* 65

66 (Modified)OVSForwardingModel Ingress* Packet* Parser* MatchC Action* Tables* Packet$ Deparser$ Egress* 4 SupportsbothediLngmodes: 4 Inline%EdiTng% 4 PostYpipeline%EdiTng% 66

67 NaïveMappingfromP4toOVS AnaïvecompilaLonofL2L3YACL%benchmarkapplicaLon PISCES(unopLmized) OVS Throughput(Gbps) Performanceoverheadof ~40% PacketSize(Bytes) 67

68 CausesofPerformanceDegradaLon Ingress* Packet* Parser* MatchC Action* Pipeline* Packet* Deparser * Egress* CPU%Cycles%per%Packet% 68

69 CausesofPerformanceDegradaLon 4 FactorsaffecLngCPUcycles: 4 Extra%copy%of%headersinthepost4pipelineediLng mode 4 FullyYspecified%checksumcalculaLon 4 Redundant%parsingofheaderfieldsandmore 69

70 CausesofPerformanceDegradaLon Factor#1:Extra%copy%of%headers% EdiTng%Mode% Pros% Cons% Post4Pipeline Extra%copy%of%headers% Inline No%extra%copy%of%headers% 4 PostYpipelineediLngconsumes2xmorecyclesthaninline edilngwhenparsing%vxlan%protocol.

71 CausesofPerformanceDegradaLon Factor#1:Extra%copy%of%headers% EdiTng%Mode% Pros% Cons% Post4Pipeline Packets%are%adjusted%once% Extracopyofheaders Inline Noextracopyofheaders MulTple%adjustments%to%packet% CyclesperPacket Post4PipelineEdiLng InlineEdiLng Inline%ediTng% PostYpipeline%ediTng% Deparse x1 x2 x4 x8 x16 Numberofadjustments

72 CausesofPerformanceDegradaLon Factor#2:FullyYSpecified%Checksums% Checksum$(* ****version,*ihl,*diffserv,*totallen,** ****identification,*flags,*fragoffset,** ****ttl,*protocol,*hdrchecksum,******* ****srcaddr,*dstaddr)* Incremental8Checksum$(ttl)* Ingress* Packet* Parser* decrement(ttl)* Packet* Deparser * Egress*

73 CausesofPerformanceDegradaLon Factor#3:Redundant%parsing%of%headers% L2 L3 L4 L2 Ingress* Packet* Parser* MatchC Action* Pipeline* Packet* Deparser * Egress*

74 OpLmizingforCPUCycles OpTmizaTons% Inlinevs.post4pipelineediLng Incrementalchecksum ParserspecializaLon AcLonspecializaLon AcLoncoalescing 74

75 OpLmizingforCPUCycles OpTmizaTons% Inlinevs.post4pipelineediLng Incrementalchecksum ParserspecializaLon AcLonspecializaLon AcLoncoalescing ExtraCopyofHeaders Fully4SpecifiedChecksum RedundantParsing 75

76 OpLmizingforCPUCycles OpTmizaTons% Inlinevs.post4pipelineediLng Incrementalchecksum ParserspecializaLon AcLonspecializaLon AcLoncoalescing 76

77 OpLmizedMappingfromP4toOVS All%opTmizaTonstogether 50 PISCES(unopLmized) PISCES(OpLmized) OVS Throughput(Gbps) Performanceoverheadof <2% PacketSize(Bytes) 77

78 Next%Steps 4 Supportforstateful%memories%andINT% % 4 IntegraTonwiththemainline%OVS% 78

79 Summary 4 SDNbroughthugechangestohownetworksoperate 4 ThereissLllamissinggapbetweenthecurrentstate ofsdnandafullyprogrammablenetwork 4 P4isanewtoolfillinthemissinggap 4 PISCES%isfirsttoshowthatprogrammabilityusingP4 comesataverysmalloverheadwhilebringinghuge benefits% 79

80 QuesLons? 80