WADS-DSN 2004 Invited Talk From Dependable Architectures ... · WADS-DSN 2004 Invited Talk From...

WADS-DSN 2004 Invited Talk

FromDependable Architectures

ToDependable Systems

Nenad MedvidovicComputer Science DepartmentUniversity of Southern CaliforniaLos Angeles, U.S.A.neno@usc.edu

Goals of the Talk

Identify some challengesSuggest some solutionsMotivate future researchInvite dissenting opinions

Goals of the Talk

The problem is this bigand sometimes ill-definedThe problem is this big

and sometimes ill-defined

Goals of the Talk

I will talk to youabout this much of it

Goals of the Talk

I will suggest a solutionfor this much of it

Goals of the Talk

But, hopefully, we will havethis much fun!

What Is Dependability?Degree of user confidence that the system will operate as expectedKey dimensions

AvailabilityReliabilitySecuritySafety

But alsoRepairabilityMaintainabilitySurvivabilityFault tolerance

What Is Architecture?A high-level model of a system

The system’s “blueprint”Represents system organization

DataComputationInteractionStructure

Embodies system propertiesCommunication integrity, performance, throughput, liveness, . . .Can/does it embody dependability?(how) Can those properties be transferred to the system itself?

A “Traditional” Architectural Model

What/where is the dependability?

A “Standard” Architectural Model

Segment(from ssi)

AttributeDescription(from serializati...

ResourceDescription(from serializati...

HttpCredential(from ht.. .

PlainRemoteResource(from adm...

HtmlHead(from ht...

SampleMuxHandler(from m...

Stresser(from tes...

JigsawHttpSessionContext(from servl...

HtmlStyle(from ht...

AdminContext(from admin)

AuthUserPrincipal(from a...

IPMatcherNode(from au...

HttpBag(from ht...

HtmlGenerator(from ht...

DirectoryResource(from resourc...

FileResource(from resourc...

HttpEntityTag(from ht...

HttpManager(from ht...

CGIHeaderHolder(from fram...

Shuffler(from ht...

IndexersCatalog(from index...

ResourceIndexer(from index...

MuxClientFactory(from m.. .

ResourceContext(from resourc...

EventManager(from time...

ResourceStoreManager(from sto...PropertySet

(from conf...

FramedResource(from resourc...

ServerHandlerManager(from daem...

AdminWriter(from adm...

Serializer(from serializati...

RequestTimeout(from ht...

MimeType(from mime)

IPMatcher(from auth)

ExternalContainer(from resourc. ..

PICS(from pi...

SSIFrame(from s...

SSIStream(from s...

ArrayDictionary(from ut...

CommandRegistry(from comman...

VariantState(from NegotiatedFra...

SampleResourceIndexer(from index...

ProtocolFrame(from frames)

MimeClientFactory(from ht...

MimeParser(from mi...

DAVMimeClientFactory(from webd...

ResourceException(from resourc...

SocketOutputBuffer(from sock...

DebugThread(from sock...

ThreadCache(from ut...

SocketClientFactoryStats(from sock...

httpd(from http)

MuxSession(from m...

ResourceFilter(from resources)

HttpMimeType(from ht...

HttpTokenList(from ht...

ResourceReference(from resources)

HTTPPrincipal(from a...

ProxyRequestObserver(from pro...

ProfileRef(from cc...

HttpExtList(from ht...

Client(from http)

CCPPRequest(from cc...

: WebBrowser : SocketClient : httpd : DirectoryResource : ContainerResource : FramedResource

: LookupResult

processRequest(Request) perform(RequestInterface) <<create>>

lookup(LookupState, LookupResult)

[lookup(LookupState, LookupResult)] [lookup(LookupState, LookupResult)]

setTarget(resourceReference)

internalLookup()

lookup(LookupState, LookupResult) [moreComponents]:

HTTPD Resources

Figure 10: Sequence diagram for the request processing use case.

: WebBrowser : SocketClient : httpd : DirectoryResource : ContainerResource : FramedResource

: LookupResult

processRequest(Request) perform(RequestInterface) <<create>>

lookup(LookupState, LookupResult)

[lookup(LookupState, LookupResult)] [lookup(LookupState, LookupResult)]

setTarget(resourceReference)

internalLookup()

lookup(LookupState, LookupResult) [moreComponents]:

HTTPD Resources

Figure 10: Sequence diagram for the request processing use case. What/where is the dependability?

But, We Can Model AnythingMeta-H, ROOM, UniCon, etc. can help ensure real-time properties in modelsMarkov chains can help ensure reliability in modelsMulti-versioning connectors can help ensure fault toleranceCode/data mobility and replication formalisms can help ensure availability. . .

So then the problem is solved, right?

Why the Problem Isn’t Solved

Why the Problem Isn’t SolvedArchitecture

Architecture(ADL)

Design(UML)

Architecture(ADL)

Implementation(middleware, bus

technolgies)

Design(UML)

Object-Oriented

Programming Language (OOPL)D/COM

CORBAC++

Visual Basic

Java Beans

Problem SpaceDomain Entities

Class Name

Attributes

Operation

Class Name

Attributes

Operation

Class Name

Attributes

Operation

Class Name

Attributes

OperationClass Name

Attributes

Operation

Architecture Space

Design Space

ComponentsConnectors

ClassesCOTS Components

Class Name

Attributes

Operation

Class Name

Attributes

Operation

Class Name

Attributes

Operation

Class Name

Attributes

OperationClass Name

Attributes

Operation

Architecture Space

Design Space

ΩWhy the Problem Isn’t Solved

Class Name

Attributes

Operation

Class Name

Attributes

Operation

Class Name

Attributes

Operation

Class Name

Attributes

OperationClass Name

Attributes

Operation

Architecture Space

Design Space

ΩΩ2 Ω3

Class Name

Attributes

Operation

Class Name

Attributes

Operation

Class Name

Attributes

Operation

Class Name

Attributes

OperationClass Name

Attributes

Operation

Architecture Space

Design Space

ΩΩ2 Ω3

Ω11Ω21Ω31Ω12 Ω22

Ω13Θ

From Models to Systems

Weather Repository

WeatherAnalyzer

ResourceManager

StrategyAnalysisKB

SimulationAgent

ResourceMonitor

SAKBUI

StrategyAnalyzer

Agent DeploymentAdvisor

Commander UI

SoldierUI

Commander UI

SoldierUI

Host 2Host 1

Host 3 Host 4 Host 5

Weather Repository

WeatherAnalyzer

ResourceManager

StrategyAnalysisKB

SimulationAgent

ResourceMonitor

SAKBUI

StrategyAnalyzer

Commander UI

SoldierUI

Commander UI

SoldierUI

The remainder of this talk will focus on two key questions:

1. How do we get from

Weather Repository

WeatherAnalyzer

ResourceManager

StrategyAnalysisKB

SimulationAgent

ResourceMonitor

SAKBUI

StrategyAnalyzer

Commander UI

SoldierUI

Commander UI

SoldierUI

Host 2Host 1

Weather Repository

WeatherAnalyzer

ResourceManager

StrategyAnalysisKB

SimulationAgent

ResourceMonitor

SAKBUI

StrategyAnalyzer

Commander UI

SoldierUI

Commander UI

SoldierUI

Host 2Host 1

Weather Repository

WeatherAnalyzer

ResourceManager

StrategyAnalysisKB

SimulationAgent

ResourceMonitor

SAKBUI

StrategyAnalyzer

Commander UI

SoldierUI

Commander UI

SoldierUI

Host 2Host 1

Weather Repository

WeatherAnalyzer

ResourceManager

StrategyAnalysisKB

SimulationAgent

ResourceMonitor

SAKBUI

StrategyAnalyzer

Commander UI

SoldierUI

Commander UI

SoldierUI

Host 2Host 1

Weather Repository

WeatherAnalyzer

ResourceManager

StrategyAnalysisKB

SimulationAgent

ResourceMonitor

SAKBUI

StrategyAnalyzer

Commander UI

SoldierUI

Commander UI

SoldierUI

Host 2Host 1

2. How do we know

?is “better” than

OutlineFrom architectures to systemsEnsuring dependability

Problem definitionProposed solution

Concluding remarks

How Do I Dependably Implement an Architecture?

Architectures provide high-level conceptsComponents, connectors, ports, events, configurations

Programming languages provide low-level constructsVariables, arrays, pointers, procedures, objects

Bridging the two often is an art-formMiddleware can help “split the difference”

Existing middleware technologiesSupport some architectural concepts (e.g., components, events)but not others (e.g., connectors, configurations)Impose particular architectural styles

How Do I Dependably Implement an Architecture?

Architectures provide high-level conceptsComponents, connectors, ports, events, configurations

Programming languages provide low-level constructsVariables, arrays, pointers, procedures, objects

Bridging the two often is an art-formMiddleware can help “split the difference”

Existing middleware technologiesSupport some architectural concepts (e.g., components, events)but not others (e.g., connectors, configurations)Impose particular architectural styles

What is needed is “architectural middleware”

Architectural MiddlewareNatively support architectural concepts as middleware constructsInclude system design support

Typically via an accompanying ADL and analysis toolsMay support explicit architectural styles

Support round-trip developmentFrom architecture to implementation and back

Support automated transformation of architectural models to implementations

i.e., dependable implementationExamples

ArchJavaAurac2.fwPrism-MW

Dependable ImplementationArchitecture

technolgies)

Design(UML)

Object-Oriented

CORBAC++

Visual Basic

Java Beans

technolgies)Object-Oriented

CORBAC++

Visual Basic

Java Beans

technolgies)Object-Oriented

CORBAC++

Visual Basic

Java Beans

IComponentIConnector

Scaffold

AbstractDispatcher

Round RobinDispatcher

AbstractScheduler

FifoScheduler

Architecture

ExtensibleComponent

Component

Connector

Serializable

IArchitecture

#mutualPort

Example: Prism-MW

Scaffold

AbstractDispatcher

AbstractScheduler

FifoScheduler

Architecture

ExtensibleComponent

Component

Connector

Serializable

IArchitecture

#mutualPort

Example: Prism-MW

Scaffold

AbstractDispatcher

AbstractScheduler

FifoScheduler

Architecture

ExtensibleComponent

Component

Connector

Serializable

IArchitecture

#mutualPort

Example: Prism-MW

Scaffold

AbstractDispatcher

AbstractScheduler

FifoScheduler

Architecture

ExtensibleComponent

Component

Connector

Serializable

IArchitecture

#mutualPort

Example: Prism-MW

Scaffold

AbstractDispatcher

AbstractScheduler

FifoScheduler

Architecture

ExtensibleComponent

Component

Connector

Serializable

IArchitecture

#mutualPort

Example: Prism-MW

Scaffold

AbstractDispatcher

AbstractScheduler

FifoScheduler

Architecture

ExtensibleComponent

Component

Connector

Serializable

IArchitecture

#mutualPort

Example: Prism-MW

Using Prism-MW

Architecture - DEMO

class DemoArch static public void main(String argv[])

Architecture arch = new Architecture ("DEMO");

Using Prism-MW

Architecture - DEMO

Architecture arch = new Architecture ("DEMO");// create componentsComponentA a = new ComponentA ("A");ComponentB b = new ComponentB ("B");ComponentD d = new ComponentD ("D");

Component BComponent A Component D

Using Prism-MW

Architecture - DEMO

Architecture arch = new Architecture ("DEMO");// create componentsComponentA a = new ComponentA ("A");ComponentB b = new ComponentB ("B");ComponentD d = new ComponentD ("D");// create connectorsConnector conn = new Connector("C");

Component BComponent A Component D CConnector C

Using Prism-MW

Architecture - DEMO

Architecture arch = new Architecture ("DEMO");// create componentsComponentA a = new ComponentA ("A");ComponentB b = new ComponentB ("B");ComponentD d = new ComponentD ("D");// create connectorsConnector conn = new Connector("C");// add components and connectors arch.addComponent(a);arch.addComponent(b);arch.addComponent(d);arch.addConnector(conn);

Component BComponent A

Component D

CConnector C

Using Prism-MW

Architecture - DEMO

Architecture arch = new Architecture ("DEMO");// create componentsComponentA a = new ComponentA ("A");ComponentB b = new ComponentB ("B");ComponentD d = new ComponentD ("D");// create connectorsConnector conn = new Connector("C");// add components and connectors arch.addComponent(a);arch.addComponent(b);arch.addComponent(d);arch.addConnector(conn);

Component D

CConnector C

// establish the interconnectionsarch.weld(a, conn);arch.weld(b, conn);arch.weld(conn, d)

Deploying a Prism-MW Architecture

add (DataRepository

weld (TopDistributionConnector,

: source HQ) : HQ;

C1_AvailableTroops) : Commander1;

Concluding remarks

AvailabilityThe degree to which a system is operational and accessible when required for use [IEEE]Deployment architecture influences availability

Components on the same host can communicate regardless of the network’s statusComponents on different hosts are insulated from each other’s failures

Quantifying availabilityRatio of the

number of successfully completed interactionsin the system to the

total number of attempted interactions

Maximizing Availability a prioriWe may not know many relevant system parameters

Dependability of each componentFrequency of component interactionsVolume of component interactionsDependability of component interactionsCPU usage on each hostDependability of each hostEffective bandwidth of each network connectionDependability of each network connection. . .

Maximizing Availability a prioriWe may not know many relevant system parameters

Dependability of each componentFrequency of component interactionsVolume of component interactionsDependability of component interactionsCPU usage on each hostDependability of each hostEffective bandwidth of each network connectionDependability of each network connection. . .

The current deployment architecture may not work well

Simplified Problem Definition

(1)a set C of n components ( Cn = ), a relation ℜ→×CCfreq : , and a function ℜ→Cmemcomp :

Given:

⎭⎬⎫

⎩⎨⎧

jiji ccifcandcbetweencommoffrequency

ccifccfreq

⎭⎬⎫

⎩⎨⎧

jiji ccifcandcbetweencommoffrequency

ccifccfreq

cformemoryrequiredcmemcomp =)(

Given:

(2) a set H of k hardware nodes ( Hk = ), a relation

ℜ→× HHrel : , and a function ℜ→Hmemhost :

Simplified Problem DefinitionGiven:

⎪⎭

⎪⎬

⎪⎩

⎪⎨

hhifhandhbetweenlinktheofyreliabilithtoconnectednotishif

hhifhhrel 0

⎪⎭

⎪⎬

⎪⎩

⎪⎨

hhifhandhbetweenlinktheofyreliabilithtoconnectednotishif

hhifhhrel 0

hhostonmemoryavailablehmemhost =)(

Simplified Problem DefinitionGiven:(3) Two relations that restrict locations

of software components 1,0: →×HCloc 1,0,1: −→×CCcolloc

⎭⎬⎫

⎩⎨⎧

jiji hontodeployedbecannotcif

hontodeployedbecancifhcloc

⎭⎬⎫

⎩⎨⎧

jiji hontodeployedbecannotcif

hontodeployedbecancifhcloc

⎪⎭

⎪⎬

⎪⎩

⎪⎨

⎧−=

candcofncollocatioonnsrestrictionoarethereifcashostsametheonbetohascifcashostsametheonbecannotcif

cccolloc011

Find a function HCf →: such that the

∑∑

∗= n

ccfreq

cfcfrelccfreqA

))(),((),(

system’s overall availability

is maximized, and the following three conditions are satisfied:

∑∑

∗= n

ccfreq

cfcfrelccfreqA

))(),((),(

Simplified Problem DefinitionFind a function HCf →: such that the

∑∑

∗= n

ccfreq

cfcfrelccfreqA

))(),((),(

⎪⎭

⎪⎬⎫

⎪⎩

⎪⎨⎧

≤=∈∀∈∀ ∑j

ihostjcompij hmemcmemhcfnjki )())()(],1[],1[

∑∑

∗= n

ccfreq

cfcfrelccfreqA

))(),((),(

1))(,(],1[ =∈∀ jj cfclocnj

∑∑

∗= n

ccfreq

cfcfrelccfreqA

))(),((),(

],1[],1[ nlnk ∈∀∈∀))()(()1),(( lklk cfcfcccolloc =⇒=))()(()1),(( lklk cfcfcccolloc ≠⇒−=

Simplified Problem DefinitionFind a function HCf →: such that the

∑∑

∗= n

ccfreq

cfcfrelccfreqA

))(),((),(

⎪⎭

⎪⎬⎫

⎪⎩

⎪⎨⎧

≤=∈∀∈∀ ∑j

ihostjcompij hmemcmemhcfnjki )())()(],1[],1[ 1))(,(],1[ =∈∀ jj cfclocnj],1[],1[ nlnk ∈∀∈∀

))()(()1),(( lklk cfcfcccolloc =⇒=))()(()1),(( lklk cfcfcccolloc ≠⇒−=

Note that the possible number of

different functions is

Concluding remarks

Overview of the ApproachObjective

Identify the problemLog and examine system eventsActively monitor the system during runtime

Develop a solutionDecide which data to cacheDecide which components to replicateIntroduce multiple execution modesCalculate an improved system deployment

Apply the solution to eliminate the problemCache or hoard data Replicate data or codeRedeploy (parts of) the system

Overview of the ApproachObjective

Identify the problemLog and examine system eventsActively monitor the system during runtime

Develop a solutionDecide which data to cacheDecide which components to replicateIntroduce multiple execution modesCalculate an improved system deployment

Apply the solution to eliminate the problemCache or hoard data Replicate data or codeRedeploy (parts of) the system

Improving Availability via Redeployment

Availability

TM TRT

TOTET’M T’R

T’T’OT’E

First Identify the Problem

Availability

TM TRT

First Identify the Problem

Availability

TM TRT

AbstractMonitor

Scaffold

AbstractDispatcher

AbstractScheduler

FifoScheduler

Architecture

ExtensibleComponent

Component

Connector

Serializable

IArchitecture

#mutualPort

Disconnection Rate

EvtFrequency

Monitoring in Prism-MW

Monitoring in Action

29 1 28

Then Develop a Solution

Availability

TM TRT

Then Develop a Solution

Availability

TM TRT

1000000

Time taken(in ms)

Achieved availability

10 comps 100 comps 200 comps 1000 comps 100 comps 30 comps 300 comps 4 hosts 10 hosts 20 hosts 100 hosts 40 hosts 7 hosts 70 hosts

Suite of algorithms:Exact – exponential complexityStochastic – quadratic complexityAdaptive greedy – cubic complexityDecentralized – quadratic complexity

Estimating in Action

Automatic Algorithm Selection

T0STES

GreedyAG

AEExact

TEG TEE

T0G T0E

Availability

TRTR TRT

Stochastic

Automatic Algorithm Selection

T0STES

GreedyAG

AEExact

TEG TEE

T0G T0E

Availability

TRTR TRT

Stochastic

Then Apply the Solution

Availability

TM TRT

Then Apply the Solution

Availability

TM TRT

Scaffold

AbstractDispatcher

AbstractScheduler

FifoScheduler

Architecture

ExtensibleComponent

Component

Connector

Serializable

IArchitecture

#mutualPort

AbstractAdmin

DeployerAbstract

Redeployment Algo

Stochastic

Adaptive Greedy

Redeployment in Prism-MW

Redeployment in Action

0Admin

Deployer

Goingfrom

Host 2Host 1

Weather Repository

WeatherAnalyzer

ResourceManager

StrategyAnalysisKB

SimulationAgent

ResourceMonitor

SAKBUI

StrategyAnalyzer

Commander UI

SoldierUI

Legend:

User interface component(fixed to a host)

Processing or datacomponent

Commander UI

SoldierUI

Host 2

Hardware host

Network link

Interaction pathbetween components

Concluding remarks

Concluding RemarksStill so much to do

Enriching/completing the modelsFocusing on additional aspects of dependabilityAddressing feature interactionsAddressing emergent propertiesDetermining which concerns are (not) architectural

Promise or Illusion?If we can define it, we should be able to

Analyze for itBuild itMeasure itAct on it

So, why haven’t we done it yet?

The Playing FieldDependability

HardwareProperties

SoftwareProperties

Availability

Reliability

Security

Safety

InteractionFrequency

Interaction Volume

Component Size

Interaction Volume

ComponentCPU Usage

AvailableMemory

Available CPU

NetworkReliability

NetworkBandwidth

Survivability

ComponentFailure Rate

H/W Properties

S/WProperties

Dependability

Exploring the Problem Space

Questions?

Using Prism-MW

Component B handles the event and sends a response

public void handle(Event e)

if (e.equals(“Event_D”)) ... Event e1= new Event(“Response_to_D”);e1.addParameter("response", resp);send(e1);...

Architecture - DEMO

Component D

CConnector C

Component D sends an event

Event e = new Event (“Event_D”);e.addParameter("param_1", p1);send (e);

Assumptions

Availability

TM TRT

TM+TE+TR<<T0TM+TE+TR<<T0

•Possible in small amounts of time even for slow links

For example, •dialup (56K), 50% reliability, 100K in 28 sec•1M, 50% reliability, 100K in less than 0.7 sec

Assumptions

Availability

TM+TE+TR<<T0TM+TE+TR<<T0

Average value ofa parameter

MonitoringHow many samples before I assume stability?Proposed solution

Average for every new event that arrivesCompare averages for small number of samples (e.g., 3)As soon as the difference between the samples is < εassume the parameter is stableSome parameters will take longer to stabilize

Keep comparing all the parameters to ensure stability, until allof them are within their ε

Keep monitoring during redeployment estimation to make sure the parameters have not changed

What Is Needed?

Scalable and lightweight support for distributed architectureswith arbitrary topologies

Support for automatic monitoringEfficient and precise solutions for the exponentially complex redeployment problemSupport for automatic redeployment

Problem breakdown

Problem breakdown1) Lack of knowledge about runtime system parameters

unknown at the time of initial deploymentchange at runtimearchitectural middleware with monitoring support

2) Exponentially complex problemn components and k hosts = kn possible deploymentssystem constraints further complicate the problempolynomial-time approximating algorithms

3) Assessing deployment architecturescomparison of different algorithmsperformance vs. complexity, sensitivity analysis, etc.analysis and simulation environment

4) Effecting the selected solutionredeploying componentsrequires an automated solutionarchitectural middleware with deployment support

Architecture(ADL)

technolgies)

Design(UML)

Object-Oriented

CORBAC++

Visual Basic

Java Beans

Not Nece

ssarily

to Sca

leWhy the Problem Isn’t Solved

WADS-DSN 2004 Invited Talk From Dependable Architectures ... · WADS-DSN 2004 Invited Talk From...

Documents

DSN Telephone Directory Global DSN Operator: (312)-560-1110

WADS News Rosh Hashanah 2014

Best Induction sealing wads Manufacturers in India - Kapil Packaging

Chap 05 dsn

Optimal Möbius Transformation for Information Visualization and …eppstein/pubs/BerEpp-WADS-01-omt-slides.… · Optimal Möbius Transformation D. Eppstein, UC Irvine, WADS 2001

NEODeS - dsn-info.fr · 2018-03-01 1 / 265 . NEODeS Cahier technique DSN. Version : CT2019.1.1 Cahier Technique DSN. Rédaction : Caisse nationale d’assurance vieillesse

Premier Holotech, Bengaluru, Hologram Wads

WADS 20ga, 28ga and .410cal. - feskent.co.ukfeskent.co.uk/wads2.pdf · WADS 20ga, 28ga and .410cal. 20ga B&P Plastic Wads. Type Case Lead Shot * Per 100 Per 500 Per 1000 B&P Z2M-15

cahier Technique De La Dsn Phase 3 P3.2017.1 - Dsn · 2017-01-09 1 / 236 NEODeS Cahier technique DSN Phase 3 Version : 2017.P3.1 Cahier Technique DSN phase 3 Rédaction : Caisse nationale

wads.org WADS 2017 - Carleton Universitypeople.scs.carleton.ca/~wads/CallForPapers/WADScfp.… · · 2016-12-24Vangelis%Markakis,!Athens!University!of!Economics!and!Business,!Greece!

RADIANCE @ DSN 2016

CHAPTER VIII LIFE AT CAMP WADS WORTH VIII LIFE AT CAMP WADS WORTH ... Two very well-known colleges are located in Spartanburg. ... which certified that Mr. Blank,

The DSN Array Development Program - NASA...The DSN Array Development Program Sander Weinreb, 818-354-4065 May 23, 2002 •Why Arrays for DSN? ... Tsys A/Tsys Year Finished DSN 70m

DSN 50th Celebration Symposium DSN : Past, Present, and Future

Best Manufacturer of Induction Sealing Wads for Glass Bottles in Delhi

WADS News - Issue 21 - Spring 2014

Dsn Optimization Hussain

Pi, WADS - IRC

The 46th Annual IEEE/IFIP International Conference on ...dsn-2016.sciencesconf.org/conference/dsn-2016/DSN2016_Final... · The 46th Annual IEEE/IFIP International Conference on Dependable

The Parametric Closure Problem - ics.uci.edueppstein/pubs/Epp-WADS-15-slides.pdf · The Parametric Closure Problem David Eppstein 14th Algorithms and Data Structures Symp. (WADS 2015)