Tiramisu: Fast Multilayer Network Verification...Tiramisu: Fast Multilayer Network Verification...

Tiramisu: Fast Multilayer Network Verification

Anubhavnidhi “Archie” Abhashkumar*, Aaron Gember Jacobson#,and Aditya Akella*

*University of Wisconsin-Madison, # Colgate University

Distributed Control Planes

Policies

Y à Z

Configuration Exampleinterface GigabitEthernet0/1

ip address 1.0.1.1 255.255.0.0ip ospf cost 1

router ospf 10 network 3.0.1.2 0.0.255.255 area 0

Policies

Y à Z

Policies

Y à Z

Policies

Y à Z

OSPFOSPF

OSPFOSPF 1

Policies

Y à Z

OSPFOSPF

OSPFOSPF 1

Router Configurations are Complex

interface GigabitEthernet0/1ip address 1.0.1.1 255.255.0.0ip ospf cost 1!interface GigabitEthernet0/2ip address 10.11.11.1 255.255.0.0!router bgp 300neighbor 2.2.2.2 route-map COMM out!route-map COMM permit 10 set community 1:1 additiveset local-preference 150!router ospf 10 network 3.0.1.2 0.0.255.255 area 0

• Multiple routing protocols• BGP• OSPF• …

• Multiple routing metrics• ospf cost• local preference• …

• Multiple filters• Community• …

Configuration Complexity Make Errors Common

Policy violations manifest under failures

Configuration Complexity Make Errors Common

Verification tools can proactively check for failures

Multiple Network Verification tools

13Performance

Bagpipe(OOPSLA’16)

Plankton (NSDI’20)

Minesweeper (SIGCOMM ’17)

ERA (OSDI’16)

Batfish (NSDI’15)

ARC (SIGCOMM’16)

P-REX (CONEXT’18)

14Performance

ERA (OSDI’16)

Batfish (NSDI’15)

ARC (SIGCOMM’16)

P-REX (CONEXT’18)

Goal: A verification tool that has good coverage and good performance

Tiramisu (NSDI’20)

15Performance

ERA (OSDI’16)

Batfish (NSDI’15)

ARC (SIGCOMM’16)

P-REX (CONEXT’18)

Goal: A verification tool that has good coverage and good performance

Tiramisu (NSDI’20)

Performance VS Coverage• ARC (SIGCOMM’16)• Graph algorithms

Network configurations

Weighted Graphs

Performance VS Coverage• ARC (SIGCOMM’16)• Graph algorithms

Weighted Graphs

• Minesweeper (SIGCOMM’17)• Symbolic encoding

Routing Algorithms

Policies

SMT Constraints

Z3 solver

Cross Layer Dependency

eBGP eBGP

OSPFOSPF

eBGP eBGP

OSPFOSPF

IBGP uses OSPF computed route to reach next hop router

eBGP eBGP

OSPFOSPF

Dst NextHop

ROUTER C

ROUTER B

ROUTER D

eBGP eBGP

OSPFOSPF

Dst NextHop

ROUTER C

ROUTER B

ROUTER D

C à E

eBGP eBGP

OSPFOSPF

Dst NextHop

ROUTER C

ROUTER B

ROUTER D

C à E

eBGP eBGP

OSPFOSPF

Dst NextHop

ROUTER C

ROUTER B

ROUTER D

C à E

Cross Layer Dependency: iBGP - OSPF

eBGP eBGP

OSPFOSPF

Dst NextHop

ROUTER C

ROUTER B

ROUTER D

C à E

eBGP eBGP

OSPFOSPF

Dst NextHop

ROUTER C

ROUTER B

ROUTER D

C à E

eBGP eBGP

OSPFOSPF

Dst NextHop

ROUTER C

ROUTER B

ROUTER D

eBGP eBGP

OSPFOSPF

Dst NextHop

ROUTER C

ROUTER B

ROUTER DDst NextHop

eBGP eBGP

OSPFOSPF

Dst NextHop

ROUTER C

ROUTER B

ROUTER DDst NextHop

ARC and Minesweeper

• ARC (SIGCOMM’16)

Ebgp Bbgp

Cospf Cospf

• Insufficient feature coverage• No IBGP, local preference,

community, ….

ARC and Minesweeper• Minesweeper (SIGCOMM’17)

• ARC (SIGCOMM’16)

Ebgp Bbgp

Cospf Cospf…….192.0.0.0 ≤ out.prefix out.prefix ≤ 192.1.0.0out.valid ⇒ truebest.valid ⇒ out.lp = 120best.valid ⇒ out.ad = 20…….

…….192.0.0.0 ≤ out.prefix out.prefix ≤ 192.1.0.0out.valid ⇒ truebest.valid ⇒ out.lp = 120best.valid ⇒ out.ad = 20…….

• Insufficient feature coverage• No IBGP, local preference,

community, ….

• Poor performance• replicate model for iBGP

InsightsConfigurations

NETWORK MODEL

VERIFICATIONALGORITHMS

NETWORK MODEL

Model VerificationARC Graph Graph

Minesweeper SMT SMT

NETWORK MODEL

Minesweeper SMT SMT

low coverage

high coverage low performance

high performance

NETWORK MODEL

Minesweeper SMT SMT

low coverage

high coverage low performance

high performance

Insight 1: Decouple network encoding from verification algorithm

NETWORK MODEL

Multilayer graph with vector of edge weights

NETWORK MODEL

VERIFICATIONALGORITHM-1

VERIFICATIONALGORITHM-N….

Categories policies based on fidelity

NETWORK MODEL

OSPFOSPF

S1Multilayer graph with vector of edge weights

NETWORK MODEL

OSPFOSPF

Policy 1: Which path is preferred?C→B << C→D→B

NETWORK MODEL

OSPFOSPF

C→B : ospf cost = 1C→D→B : ospf cost = 6

NETWORK MODEL

OSPFOSPF

PATHENUMERATION C→B : ospf cost = 1

C→D→B : ospf cost = 6

NETWORK MODEL

OSPFOSPF

PATHENUMERATION

Policy 2: Can C reach B with 1 link failure?

NETWORK MODEL

OSPFOSPF

PATHENUMERATION

NETWORK MODEL

OSPFOSPF

PATHENUMERATION

Min-cut = 2

NETWORK MODEL

OSPFOSPF

PATHENUMERATION

QUANTITATIVE GRAPH

PROPERTY Min-cut = 2

NETWORK MODEL

OSPFOSPF

PATHENUMERATION

QUANTITATIVE GRAPH

PROPERTY

Policy 3: Is C always unreachable/blocked from B?

NETWORK MODEL

VERIFICATIONALGORITHM-N

OSPFOSPF

PATHENUMERATION

QUANTITATIVE GRAPH

PROPERTY

Policy 3: Is C always unreachable/blocked from B?

PATHEXISTENCE

(CONNECTIVITY)

NETWORK MODEL

VERIFICATIONALGORITHM-N

OSPFOSPF

PATHENUMERATION

QUANTITATIVE GRAPH

PROPERTY

PATHEXISTENCE

(CONNECTIVITY)

low performancehigh fidelity

high performancelow fidelity

Insight 2: Different properties require different levels of fidelity modeling of the control plane. Use property-specific algorithm for performance benefits

Tiramisu Overview

Traffic Propagation Graph (TPG)

Routing Adjacencies Graph (RAG)

ConfigurationsGraph Model

Tiramisu Overview

Configurations

PATHENUMERATION

Graph Model

Algorithms

Tiramisu Overview

Configurations

PATHENUMERATION

NUMERIC GRAPH PROPERTY

Graph Model

Algorithms

ILPTPVP

Tiramisu Overview

Configurations

PATHENUMERATION

PATHEXISTENCE

Graph Model

Algorithms

ILPTPVP

Tiramisu Overview

ConfigurationsGraph Model

iBGPeBGP eBGP

OSPFOSPF1

• Vertices: RIB of a routing processes and ingress/egress point of a switch/router

Bvlan:BC:in

Bvlan:BD:inEvlan:BE:in

Bospf Cospf Dospf

Ebgp Bbgp Cbgp Dbgp

Evlan:BE:out Bvlan:BD:out

Bvlan:BE:in Cvlan:BC:in

Cvlan:BC:out

Cvlan:CD:in

Cvlan:CD:out

Dvlan:CD:in

Dvlan:CD:out

Dvlan:BD:in

Bvlan:BC:out Dvlan:BD:outBvlan:BE:out

iBGPeBGP eBGP

OSPFOSPF1

• Vertices: RIB of a routing processes and ingress/egress point of a switch/router• Edges: establish route dependency and traffic flow

Bvlan:BC:in

Bospf Cospf Dospf

Ebgp Bbgp Cbgp Dbgp

Cvlan:BC:out

Cvlan:CD:in

Cvlan:CD:out

Dvlan:CD:in

Dvlan:CD:out

Dvlan:BD:in

iBGPeBGP eBGP

OSPFOSPF1

• Vertices: RIB of a routing processes and ingress/egress point of a switch/router• Edges: establish route dependency and traffic flow

Bvlan:BC:in

Bospf Cospf Dospf

Ebgp Bbgp Cbgp Dbgp

Cvlan:BC:out

Cvlan:CD:in

Cvlan:CD:out

Dvlan:CD:in

Dvlan:CD:out

Dvlan:BD:in

iBGPeBGP eBGP

OSPFOSPF1

• Vertices: RIB of a routing processes and ingress/egress point of a switch/router• Edges: establish route dependency and traffic flow• Vector of edge weights: multiple route metrics

Bvlan:BC:in

Bospf Cospf Dospf

Ebgp Bbgp Cbgp Dbgp

Cvlan:BC:out

Cvlan:CD:in

Cvlan:CD:out

Dvlan:CD:in

Dvlan:CD:out

Dvlan:BD:in

Bvlan:BC:out

<ospf:1> <ospf:1>

<len:0><len:0>

<ospf:5>

<len:0><len:1>

<ospf:5> <ospf:1>Dvlan:BD:outBvlan:BE:out

iBGPeBGP eBGP

OSPFOSPF1

Tiramisu Overview

Configurations

PATHENUMERATION

Graph Model

Algorithms

TPVP - Tiramisu Path Vector Protocol• Griffin et al. (TON’2002) and Sobrinho (TON’2005) models stable paths

problem as Simple Path Vector Protocol (SPVP) and routing algebra• TPVP is derived from SPVP and is modeled on routing algebra

• ⊕ operator to model path cost computation• ⪯ operator to model preference relation and path selection

Cospf<ospf:1>

<ospf:1><len:0>Cbgp

Cospf<ospf:1>

<ospf:1><len:0>Cbgp

{ospf:5} Node ⊕ ⪯

Cospf {ospf:1} ⊕ospf {ospf:5}= {ospf:6}

{ospf:6} ⪯ ospf {}

{ospf:6} {}

Cospf<ospf:1>

<ospf:1><len:0>Cbgp

{ospf:1}

Node ⊕ ⪯

Cospf {ospf:1} ⊕ospf {ospf:1} = {ospf:2}

{ospf:2} ⪯ ospf {ospf:6}

{ospf:2} {}

Cospf<ospf:1>

<ospf:1><len:0>Cbgp {ospf:2}

{ospf:2} {len:0, ospf:2}

Node ⊕ ⪯

Cbgp {len:0} ⊕bgp {ospf:1}= {len:0, ospf: 2}

{len:0, ospf:2} ⪯ bgp {}

Cospf<ospf:1>

<ospf:1><len:0>Cbgp

{ospf:2} {len:0, ospf:2}

{ospf:5} Node ⊕ ⪯

Cospf {ospf:1} ⊕ospf {ospf:5}= {ospf: 6}

{ospf:6} ⪯ ospf {ospf:2}

TPVP - Tiramisu Path Vector Protocol

OSPFOSPF

P1: C→D→B : ospf cost = 2P2: C→B : ospf cost = 4

OSPFOSPF

TPVP - Tiramisu Path Vector Protocol• To verify path preference (P1 << P2), Tiramisu multiple instances of TPVP for

different failure scenarios• Runs TPVP three times

OSPFOSPF

Tiramisu Overview

Configurations

Graph Model

Algorithms

ILPs - Integer Linear Programs

• Traffic flows in the direction opposite to route advertisement

eBGPeBGPS1

<at:c1>

<bt:c1>

• Traffic flows in the direction opposite to route advertisement• Only depends on quantitative path property and not exact path

min cut is 1

eBGPeBGPS1

<at:c1>

<bt:c1>B

• Traffic flows in the direction opposite to route advertisement• Only depends on quantitative path property and not exact path• ILP constraints model reachability• Tag/Community

min cut is 1

eBGPeBGPS1

<at:c1>

<bt:c1>B

• Objective models the graph property75

eBGPeBGPS1

<at:c1>

<bt:c1>B

Reachability < K failures = Minimize link failuresLongest path = Maximize path length

• Objective models the graph property• More details about the ILP and its corner cases are in the paper 76

eBGPeBGPS1

<at:c1>

<bt:c1>B

Reachability < K failures = Minimize link failuresLongest path = Maximize path length

Tiramisu Overview

Configurations

PATHEXISTENCE

Graph Model

Algorithms

TDFS – Tiramisu Depth First Search

eBGPeBGPS1

<bt:c1>

<at:c1>

• Can’t use vanilla graph algorithms to model tags

eBGPeBGPS1

<bt:c1>

<at:c1>

• Can’t use vanilla graph algorithms to model tags• Conditions• Block path: tag-blocking node à tag-adding node

eBGPeBGPS1

<bt:c1><rt:c1>

<at:c1>

• Can’t use vanilla graph algorithms to model tags• Conditions• Block path: tag-blocking node à tag-adding node• Allow path: tag-blocking node à tag-removing node à tag-adding node

eBGPeBGPS1

<bt:c1><rt:c1>

<at:c1>Verify if src and dst are always unreachable

• Can’t use vanilla graph algorithms to model tags• Conditions• Block path: tag-blocking node à tag-adding node• Allow path: tag-blocking node à tag-removing node à tag-adding node

Tiramisu Overview

Configurations

PATHENUMERATION

PATHEXISTENCE

Graph Model

Algorithms

ILPTPVP

Evaluation• Networks used • Real networks: 4 universities and 34 datacenters

• Evaluation• Tiramisu verification performance• Comparison with other state-of-the-art

Policy Name Algorithm

block Always blocked TDFS

bound Always bounded length ILP

pref Path preference TPVP

Evaluation – Tiramisu’s Performance

• bound is slowest because it uses ILP

Uni1 (9) Uni2 (24) Uni3 (26) Uni4 (35)Ti

University

pref bound block

• bound is slowest because it uses ILP• block is fastest because it uses TDFS

Uni1 (9) Uni2 (24) Uni3 (26) Uni4 (35)Ti

University

pref bound block

• bound is slowest because it uses ILP• block is fastest because it uses TDFS• pref is slower than block as TPVP is more complex

Uni1 (9) Uni2 (24) Uni3 (26) Uni4 (35)Ti

University

pref bound block

• bound is slowest because it uses ILP• block is fastest because it uses TDFS• pref is slower than block as TPVP is more complex• For large networks, pref is as long as bound• Large networks à More candidate and longer paths à More calls to TPVP

Uni1 (9) Uni2 (24) Uni3 (26) Uni4 (35)Ti

University

pref bound block

Evaluation: Tiramisu vs. Minesweeper (No failures)

• block has the most speedup

0 5 10 15 20 25

Network Size

0 10 20 30

Network Size

0 10 20 30

Network Size

Evaluation: Tiramisu vs. Minesweeper (No failures)

• block has the most speedup• pref has the least speedup (especially for large networks)• Large networks à more and longer candidate paths à more calls to TPVP

0 5 10 15 20 25

Network Size

0 10 20 30

Network Size

0 10 20 30

Network Size

Evaluation: Tiramisu vs. Minesweeper (All failures)

• Same trend but significantly better speed up• Minesweeper uses same encoding for all policies• Tiramisu uses property specific algorithms

020406080

0 5 10 15 20 25

Network Size

0 10 20 30

Network Size

0 10 20 30

Network Size

Summary• Tiramisu decouples encoding of the network from verification algorithms• Tiramisu uses a multilayer graph control plane model• Tiramisu uses• TPVP to enumerate paths• ILP to measure quantitative graph property• TDFS to check path existence

• Tiramisu achieves good performance without losing too much coverage• No external advertisements

Tiramisu: Fast Multilayer Network Verification...Tiramisu: Fast Multilayer Network Verification...

Documents

Transcript of Tiramisu: Fast Multilayer Network Verification...Tiramisu: Fast Multilayer Network Verification...

Tiramisu Pentru Copii

Multilayer Varistors(Automotive Grade) Multilayer Varistor ...

Multilayer Microfluidics

Multilayer Campus Architecture and Design Principalsd2zmdbbm9feqrf.cloudfront.net/2015/usa/pdf/BRKCRS-2031.pdf · Multilayer Campus Architecture and Design Principals ... • Multilayer

Multilayer Neural Networksmlss.tuebingen.mpg.de/2013/bottou_slides.pdf · Multilayer Neural Networks (are no longer old-fashioned!) MLSS TUEBINGEN 2013 ... Training multilayer networks

Tiramisu: Fast Multilayer Network Verification › system › files › nsdi20-paper-abhash... · 2020-02-25 · Tiramisu: Fast Multilayer Network Veriﬁcation Anubhavnidhi Abhashkumar∗,

tempting tiramisu trifle - Tupperware USAorder.tupperware.com/coe-pdf/recipe_12book_tiramisu_sp.pdfThe Italian dessert, tiramisu, ... ¼ cup plus 1 tsp. Simple Indulgence Chocolate

Cisco Networking Academy CCNP – Multilayer Switching Multilayer ...

CHILISIN ELECTRONICS CORP. - Промэлектроника · 13 Comprehensive Product Portfolio Multilayer Power Inductor Multilayer Chip Inductor Multilayer Chip Bead Multilayer

Multilayer Neural Networks (sometimes called “Multilayer Perceptrons” or MLPs)

Tiramisu Thermomix

Estimation of Complex Permittivity Composite Multilayer ... · PDF fileComposite Multilayer Material at Microwave Frequency Using Waveguide Measurements ... Composite Multilayer Material

Perceptron Multilayer

Tiramisu: Fast and General Network Veriﬁcation · not model certain features of the network. In this paper, we propose a new multilayer hedge graph abstraction, Tiramisu, that supports

Italian Tiramisu Recipe Easy–How to Make Real Tiramisu Italiano Dessert

22-Theory of multilayer filmsoptics.hanyang.ac.kr/~shsong/22-Theory of multilayer... · 2016-08-31 · 22. Theory of Multilayer Films22. Theory of Multilayer Films Transfer Matrix

GCSE Food Technology – Layered Deserts: Tiramisu:

ficha tecnica tiramisu - exclusivasdiper.com

TIRAMISU: A Polyhedral Compiler for Expressing Fast and …groups.csail.mit.edu › commit › papers › 18 › tiramisu_paper.pdf · 2018-12-20 · TIRAMISU: A Polyhedral Compiler

Multilayer Varistors Multilayer Varistor for ESD ... - PSD API