These Diallo

Doctor of Science ThesisUPMC Sorbonne Universites

Specialization

COMPUTER SCIENCE

presented the 7th march 2013 by

Mr. Mohamed BOBO DIALLO

Submitted in partial satisfaction of the requirements for the degree of

DOCTOR OF SCIENCE of UPMC Sorbonne Universites

Content-based Networking forGlobal Scale Mediation Services

Commitee in charge:

Serge FDIDA Advisor Professor, UPMC Sorbonne universites { Paris 6

Promethee SPATHIS Advisor Associate Professor, UPMC Sorbonne universites { Paris 6

Walid DABBOUS Reviewer Researcher, INRIA Sophia Antipolis

Farouk KAMOUN Reviewer Professor, SESAME

Paul MOCKAPETRIS Examiner Nominum

Sebastien TIXEUIL Examiner Professor, UPMC Sorbonne universites { Paris 6

Kave SALAMATIAN Examiner Professor, Universite de Savoie

Remerciements

Je remercie Serge Fdida qui a rendu possible la redaction de ce manuscrit.Je remercie Promethee Spathis qui m'a ouvert les portes du Lip6.Ensuite, je remercie ma famille qui m'a soutenu dans les moments diciles.Puis, je remercie chacun pour son aide la plus inme. Je pense a Mubashir, Minh,

Marguerite, Konstantin, Pierre-Emmanuel, Nicolas, Vasilis, Vincent, Mehdi, Youcef, Abdouet les autres.

Louange a Dieu le ma^tre du possible.

5

Abstract

Online information consumers are increasingly overwhelmed by the volume of informationavailable at global scale. As such, they have been privileging mediation services deliveringonly those publications relevant to their registered information interests. This includesaggregators, syndication and alerting services. We envision an infrastructure realizingglobal scale mediation between information providers and information consumers calledthe mediation network. Realizing such infrastructure raises several requirements. First,decentralization is an essential feature for the mediation network to be sustainable. Second,information consumers should register their interests according to expressive subscriptionlanguages. Third, it is desirable that the service integrates the search, content retrieval anddissemination activities. Finally, achieving communication-eciency is important, giventhe scarcity of end-to-end bandwidth and the huge volume of information available at largescale.

Content-based networking (CBN) is an appealing technology to provide ecient dissemi-nation channels to information providers to reach information consumers in a decentralizedand loose coupled manner. It supports naturally expressive subscription languages. Ex-isting CBN proposals support an exhaustive ltering semantic, i.e. a consumer registeringits interests will receive all the corresponding matching publications. Such semantic isappropriate for a wide range of applications including distributed games, stock quote ormonitoring applications. However, for applications such as news distribution or contentsharing, the amount of relevant publications available at global scale may be overwhelmingas information consumers have limited attention span. Implementing the same exhaustiveltering semantic for these applications would result in a huge information overload andcommunication overhead.

We extend CBN by dening a new service model supporting both content retrieval anddissemination trac, but also quantifying information consumers' attention capabilities inorder to address the information overload. Secondly, we dene the e-CBN framework whichaddresses the key design issues in implementing eciently the service model. Thirdly, wedescribe and discuss the results of extensive simulations that we led in order to quantifythe gains of the framework over baseline CBN as well as quantify the gains introduced byeach of the component of the framework. Finally, we conclude the thesis with perspectivesand open research problems.

7

Key Words:

Content-based networking, publish/subscribe, content distribution.

Table of contents

1 Introduction 15

1.1 Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

1.2 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

1.3 Contributions of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2 The E-CBN framework for extreme-scale content-based networking 21

2.1 Content-based networking (CBN): Overview . . . . . . . . . . . . . . . . . . 22

2.2 e-CBN: An enhanced service model for extreme scale content-based networking 24

2.3 e-CBN: Architecture and mechanisms . . . . . . . . . . . . . . . . . . . . . . 27

2.3.1 Router model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.3.2 Caching policies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2.3.2.1 Selection and replacement policies . . . . . . . . . . . . . . 31

2.3.2.2 Caching policies . . . . . . . . . . . . . . . . . . . . . . . . 33

2.3.3 Interest forwarding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2.3.4 Dissemination strategies . . . . . . . . . . . . . . . . . . . . . . . . . 34

2.3.4.1 Pull/delayed push (PDP) . . . . . . . . . . . . . . . . . . . . 35

2.3.5 Publication forwarding . . . . . . . . . . . . . . . . . . . . . . . . . . 36

2.3.6 Forwarding optimizations . . . . . . . . . . . . . . . . . . . . . . . . 38

2.3.6.1 Congestion-aware forwarding . . . . . . . . . . . . . . . . . 38

2.3.6.2 Handling duplicate responses . . . . . . . . . . . . . . . . . 38

2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3 Evaluation 41

3.1 Workload characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.1.1 On the relevance of workload modelling . . . . . . . . . . . . . . . . 42

3.1.2 Evaluation methodology . . . . . . . . . . . . . . . . . . . . . . . . . 43

3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

3.3 Caching policies evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

11

4 Conclusion 554.1 Summary of contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564.2 Originality of the contributions . . . . . . . . . . . . . . . . . . . . . . . . . 574.3 Open Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

List of gures 61

List of tables 63

References 67

Chapter1Introduction

Contents

1.1 Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

1.2 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . 19

1.3 Contributions of the thesis . . . . . . . . . . . . . . . . . . . . . . 20

15

16 1.1. CONTEXT

1.1 Context

The reference Internet model was adequate in the early days of the web, when information

consumers' behavior consisted in browsing a repository of web documents and fetching the

rare HTML pages of interest. Nowadays, information consumers' attention span does no

longer follow the pace of relevant information generated online at global scale and most

of them are privileging aggregators, as well as services enabling personalized delivery of

information by ltering the stream of online content according to consumer preferences at

various granularity (channel, topic or content). This includes content syndication services

enabled by the RSS [Boa07] and ATOM [NS05] standards, as well as alerting services [ale],

which enable consumers to keep abreast of the latest news related to an aair, an individual

or a subject, or to be notied of updates on specic channels or feeds. Additionally, an

increasing number of users are privileging attractive services such as Miro [mir] to access

online video, which seamlessly integrates search, pre-fetching and downloading functional-

ities, and provides access to a large number of online video stores and TV channels.

Handling old and new trac patterns in the current Internet involves several layers

of indirections including DNS resolution, which may become a bottleneck, if improperly

engineered. Moreover, trac crossing the middle mile of the Internet usually suers from

signicant delays and losses [Lei09]. As a consequence, caching proxies, DNS caching and

content distribution networks (CDN) have emerged as solutions of choice in order to meet

the requirements of content-related applications at global scale.

To fulll data-intensive applications needs, the receiver-driven approach has been re-

cently introduced. This approach is in contrast with the historical Internet model which

relies on host reachability. The receiver-driven approach instead focuses on content de-

livery given predened consumer requirements resulting in the choice of stateful switches

and in-networking caching. In fact, it has been argued that extending routers with caching

resources would eliminate the redundancy existing in the Internet trac [AGA+08]. In this

context, the ow of data is implicitly steered according to receivers' registered interests,

rather than explicitly relying on the binding of the corresponding data objects to their host

location. Data objects are directly addressed through names, which may oer dierent lev-

els of expressiveness. Several architectures have emerged under the banner of information-

centric networking as a consensus label in the research community working on receiver-

driven architectures for global communications [NBEK+] [FTP09] [JST+09] [CPW11].

We believe that the Next Generation Internet (NGI) will provide a substrate supporting

several communication architectures, including TCP/IP. It will enforce isolation, as well as

interoperability depending on stakeholders policies. The NGI will involve network operators

participating in the basic substrate, and virtual operators deploying and managing global

CHAPTER 1. INTRODUCTION 17

communication networks. Experimentation platforms such as GENI [BFH+06], ONELAB

[FFM10] and openflow [MAB+08] are already implementing part of this vision. Competing

receiver-driven architectures should demonstrate strong incentives for adoption by virtual

network operators and users as well as achieve satisfying performances.

To be successful in providing eciently global scale mediation services meeting users'

expectations receiver-driven architectures should meet two additional requirements. First,

they should provide an integrated service supporting the search, retrieval, and dissemi-

nation activities and supporting expressive subscription languages and rich metadata for

content description. In fact, rich metadata are more and more used to improve search

quality [cal]. Second, they should be decentralized. Decentralization is an incontrovertible

requirement for mediation services to sustain at global scale. Existing mediation services

operate as centralized or two-tiers systems operating over clusters of thousands of com-

puters to satisfy global scale workloads and consumer requirements. Nowadays, large-scale

distributed applications are usually deployed over clusters of computers. However, cluster

sizes are often limited by practical constraints such as power supply, cooling and switch-

ing technology limitations. For instance, it has been estimated that information retrieval

systems would require at least 30 clusters of 50000 computers by 2010 to operate, which

is obviously not sustainable [BYCJ+07]. Also, information consumers are more and more

scared about the big brother scenario emerging from the situation of monopoly of major

search engines such as Google, Bing or Yahoo.

Among emerging receiver-driven architectures, content-based networking (CBN)] [cbn]

enables receivers to register their information interests according to expressive subscription

languages and be notied with matching publications. Moreover, CBN implements a decen-

tralized content-based publish/subscribe (CBPS) [CS04] communication model and provides

ecient dissemination channels for a wide range of applications such as distributed games,

stock quote or monitoring applications. This thesis focuses on extending baseline content-

based networking (CBN) with the ability to meet large scale deployments characterized by

a large number of widely spread consumers with heterogeneous requirements, the scarcity

of end-to-end bandwidth and the information overload.

Publish/subscribe communications involve subscribers or receivers registering informa-

tion interests, and publishers or senders publishing information. The purpose of a pub-

lish/subscribe system is to mediate between subscribers and publishers by guaranteeing

the timely delivery of relevant publications to receivers. Publish/subscribe communica-

tions have been categorized according to the expressiveness of the subscription languages

in channel-based, topic-based and content-based publish/subscribe.

In channel-based publish/subscribe systems, subscribers subscribe to notications on

specic feeds or channels. A channel or feed is a localized source of information typically

18 1.1. CONTEXT

identied by an URL. In common implementations of channel-based publish/subscribe sys-

tem, subscribers poll the publishers to know whether new publications are available. This

polling behavior generates much overhead for the network that have been addressed by

many work. RSS cloud is an optional sub-element of the RSS protocol's channel element

that enables realtime push notications for feeds. This is done using a cloud element al-

lowing a software service to register with a cloud which noties subscribers of updates to

the channel. RSS cloud is not limited to RSS feeds but can also be used with other feed

formats such as ATOM. Another protocol superseding the default polling behavior of RSS (or

ATOM) is the pubsubhubbub protocol [pub]. In pubsubhubbub, a subscriber initially polls

the RSS (or ATOM) feed in the conventional way, i.e. by requesting it from the feed server.

The subscriber then inspects the feed, and if it references a hub, the subscriber can sub-

scribe to the feed URL topic on that hub. The subscriber runs a server so that hubs can

directly notify it when any of its subscribed topics have updated. Publishers expose their

content as RSS (or ATOM) feeds, but with the inclusion of hub references. They post notica-

tions to those referenced hubs whenever they publish something. Thus, when a publication

event occurs, the publisher calls its hubs and the hubs call their subscribers [wik]. Also,

FeedTree [SMPD05] and Corona [RPS06] are peer-to-peer systems for distributing Web

feeds faster and with lower bandwidth requirements for publishers. Instead of polling feeds

independently, FeedTree and Corona users cooperate to share feeds updates [wik].

In topic-based publish/subscribe systems [CMTV07] [PRGK09] [BBQ+07] [RKCD01],

messages are published on abstract event channels called topics. Users interested to receive

messages published on certain topics issue subscribe requests specifying their topics of

interest. Then, the publish/subscribe infrastructure guarantees to distribute each newly

published message to all the users that have expressed in the message's topic [CMTV07].

The notion of topic is very similar to the notion of group. Consequently, subscribing to a

topic T is equivalent to becoming a member of group T and publishing a message on topic

T can be viewed as broadcasting the message among the members of group T [EFGK03].

Content-based publish/subscribe [CS04] [CRW00] [CRW03] [AT10] [GSAA04] [MSRS09]

[CCP08] provides a ner granularity, enabling subscribers to issue predicates on publica-

tion message's content. Content-based publish/subscribe is more dynamic and expressive

than topic-based publish/subscribe. The expressiveness of content-based publish/subscribe

should be traded against increased complexity. Content-based publish/subscribe solutions

operate acoording to two modes: (1) ltering, and (2) rendezvous based [BV05]. When

ltering applies, the brokers determine the next hop(s) in the delivery tree of a message

M on a hop-by-hop basis by evaluating M against the subscription forwarding table which

aggregates the set of advertised subscriptions. When the publish/subscribe system oper-

ates as a rendezvous network, it is either based on distributed hash tables (DHTs) or on

CHAPTER 1. INTRODUCTION 19

dynamic partitioning of the event space among a set of brokers. Content-based networking

is an attempt to implement ltering-based content-based publish/subscribe as a network

level communication service.

1.2 Problem statement

Existing content-based networking schemes [CRW04] [CRW06] are designed to eciently

implement an exhaustive ltering semantic (or simply exhaustive semantic), which is ap-

propriate for many notication services. For instance, in sensor network applications, a

sink or receiver subscribing to an event such as, the temperature monitored by any sensor is

larger than a given threshold, is interested in every occurrence of that event. Consequently,

it is necessary to advertise subscriptions in the mediation network using ooding mecha-

nisms in order that any event matching a set of interests be notied to the corresponding

receivers, ideally, through a spanning tree rooted at the source of that event.

However, in many application scenarios, information consumers have a limited atten-

tion span and are in most cases interested in few responses relevant to their information

interests. In fact, with the globalization of the Internet, consumers are overwhelmed by

the huge volume of information accessible. This situation which has been popularized by

the futurologist Alvin Toer [Tof84], as information overload, has fostered the emergence

of mediation services capturing consumers' attention span in order to deliver only the

estimated most relevant content to them. Several online alerting services already enable

information consumers to keep abreast of the latest news related to a topic, a personality or

a real-world event. An example of such services is Google alerts [ale], which provides an

interface allowing consumers to personalize their alerts in terms of frequency (immediately,

daily, weekly) and volume (exhaustive search or ranked search) of notications.

Content-based networking is an appealing technology for dissemination services and

implementing the exhaustive semantic in scenarios where consumers register their infor-

mation interests jointly with a quantication of their attention span is not incongruous.

In fact, one can imagine solutions where ranking functions would be executed at the edge

of the network in order to display only the most relevant publications tting consumers'

attention span. However, devising ranking functions for decentralized content-based pub-

lish/subscribe (CBPS) is questionable, since matching consists in assessing a predicate

against meta-data describing available content according to an exact semantic and as such

diering from the behavior of search engines which help users locate a needle in a haystack.

Consequently, implementing the same exhaustive semantic in presence of information over-

load, would generate more trac than necessary to satisfy consumers.

Shifting from the exhaustive semantic towards a service model quantifying consumers'

20 1.3. CONTRIBUTIONS OF THE THESIS

attention capability, opens new opportunities to reduce the amount of trac carried by

content-based networks. Reducing the amount of trac necessary to satisfy each interest,

will reduce congestion in the network and increase the probability to timely satisfy sub-

scriptions. Moreover, it is important that content-based networking technology captures

the qualitative heterogeneity of information consumers that are not interested in future

publications only, but in publications immediately available as well. This will only increase

the attractiveness of content-based networking technology, given that the larger the number

of users, the larger online mediation services will make revenue through advertising.

1.3 Contributions of the thesis

This thesis introduces the e-CBN framework which extends baseline content-based network-

ing with the ability to meet the requirements of global scale mediation services, namely

large number of widely spread information consumers, information overload, and scarcity

of end-to-end bandwidth. e-CBN denes a generic service model supporting both content

retrieval and dissemination trac, and capturing information consumers' attention capa-

bility in order to address the information overload. Additionnaly, e-CBN leverages caching

in order to increase content availability, and denes the protocols required to eciently

implement the service model (Chapter 2). A characterization of the framework under re-

alistic workload assumptions reveals that e-CBN drastically improves performances for a

wide range of congurations (Chapter 3).

Chapter2The E-CBN framework forextreme-scale content-basednetworking

Contents

2.1 Content-based networking (CBN): Overview . . . . . . . . . . . . 22

2.2 e-CBN: An enhanced service model for extreme scale content-based networking . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.3 e-CBN: Architecture and mechanisms . . . . . . . . . . . . . . . . 27

2.3.1 Router model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.3.2 Caching policies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2.3.3 Interest forwarding . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2.3.4 Dissemination strategies . . . . . . . . . . . . . . . . . . . . . . . . 34

2.3.5 Publication forwarding . . . . . . . . . . . . . . . . . . . . . . . . . 36

2.3.6 Forwarding optimizations . . . . . . . . . . . . . . . . . . . . . . . 38

2.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

21

22 2.1. CONTENT-BASED NETWORKING (CBN): OVERVIEW

This chapter describes a framework for extreme scale content-based networking charac-

terized by the heterogeneity of information consumer requirements, the information over-

load and the scarcity of end-to-end bandwidth. The framework introduces a service model

which captures the quantitative and qualitative heterogeneity of information consumer

needs, and addresses the key design issues in implementing eciently the service model at

extreme scale.

Mediation network

Interests

Publications

Broadcasters/PublishersInformation consumers

provider 2 provider 1

provider 3provider 4

Channel providers

Figure 2.1: Mediation network

2.1 Content-based networking (CBN): Overview

Content-based networking(CBN) involves three types of entities: subscribers, receivers

or information consumers, publishers or information providers and routers or brokers.

Each receiver submits its interests by sending a subscription to the network where

routers acting as proxies are responsible for returning the corresponding matching

pieces of data. The rst router to handle the subscription advertised by a receiver,

i.e. one of the proxy mentionned above, is called a home router. Note that receivers

select their home routers on the basis of trust or proximity. Publishers upload their

publications so that they can be disseminated to interested receivers. A publication

consists of a data item and a metadata description, while an interest is described

by a predicate over the metadata space. Predicates and metadata typically follow

an attribute/value schema. A publication P matches a subscription S, whenever

CHAPTER 2. THE E-CBN FRAMEWORK FOR EXTREME-SCALECONTENT-BASED NETWORKING 23

the metadata describing P matches the predicate dened by S. Routers cooperate to

eciently and reliably disseminate data items corresponding to uploaded publications

towards interested receivers.

Content-based forwarding (CBF) is the algorithm that based on the information es-

tablished by the routing algorithm, processes incoming messages to decide on which

interface an incoming message should be forwarded. That information is compiled in

the forwarding table which associates each interface to a lter combining the pred-

icates of the descendants in the dissemination tree via that interface. We dene a

lter as a compact representation of a set of predicates. Ecient data structures

for forwarding tables are mentioned in [CW03] [TK06]. Carzaniga et al. [CRW06],

describes two content-based forwarding schemes requiring a spanning tree rooted at

each sender that can be congured through shortest-path trees or reverse-path for-

warding. However, these CBF schemes are correct only if spanning trees verify the

all-pairs path symmetry property, i.e. only in the case where shortest-paths are unique

or routes between routers are symmetric. In practice, it is dicult to enforce such

properties. Consequently, the deployment of such protocols is realistic only atop a

global spanning tree.

In order to reduce the complexity of content-based forwarding protocols, which re-

quires evaluating publications against the index of advertised subscriptions at every

hop of the dissemination tree, other work have proposed to implement matching only

at the publishing brokers and switching at subsequent brokers of the disseminaton

tree on the basis of explicit forwarding directives. For instance, the DV/DRP proto-

col [HCW+06] is proposed with an optimization which consists in doing matching

only at the source of publications and appending a particular structure which identi-

es the recipients of the message. When a message has to be pushed on two or more

interfaces, the destination set structure is aected before being attached to the mes-

sage and forwarded to the destinations downstream each interface. However, DV/DRP

uses a compact bit vector data structure to address the message, whose size equals

the number of sink nodes in the system. This assumes a small number of potential

receivers and is thus not scalable. Similarly, LIPSIN [JZER+09] is a forwarding proto-

col that encodes dissemination trees in message headers as bloom lters and achieves

line-speed forwarding. However, LIPSIN requires that each link of the topology be

assigned an identier and requires either a separation of the control plane from the

forwarding plane or/and that each router has a global overview of the topology.

Content-based routing (CBR) is the distributed algorithm that collects, propagates and

assembles receivers' interests as well as topological information to the router for-

242.2. E-CBN: AN ENHANCED SERVICE MODEL FOR EXTREME SCALE

CONTENT-BASED NETWORKING

warding functions. Existing content-based routing (CBR) [CS04] [CCP08] [CRW06]

schemes are designed to support an exhaustive semantic where receivers register for

all relevant publications matching their interests. Typically, routing consists in broad-

casting subscriptions within the network in order to congure the dissemination tree

required to eciently forward publications from senders to receivers. Content-based

routing requires a broadcast layer for operation on top of arbitrary topologies, which

can be implemented through spanning trees.

2.2 e-CBN: An enhanced service model for extreme scale content-based networking

We consider a mediation network constituted of routers or brokers, involving independent

stakeholders, interconnected according to an arbitrary topology that captures the specic

features of content networks, and that provides ecient dissemination channels for infor-

mation providers to reach information consumers at global scale.

Information providers (or publishers) upload their publications to the mediation net-

work so that they can be disseminated towards interested receivers. We assume that author-

itative publishers upload publications once to the mediation network and each publication

is assigned a unique identier. The purpose of restricting publications to authoritative

publishers is to guarantee that two publications with dierent identiers embed dierent

content.

A receiver r advertises its information needs to its home router R, as a subscription

S (predicate, max, lifetime, freshness) where predicate denes its information interest,

max species the maximum amount of publications admissible by r over a period of time

lifetime, and freshness the maximum age for a relevant publication matching the infor-

mation interest. The age of a publication is the elapsed time since its initial upload in the

mediation network. S is constituted of more attributes which are introduced throughout

the chapter.

A publication P matches a subscription S, whenever the metadata describing P matches

the predicate dened by S and the age of P is less than freshness at the end of the service

period. Routers cooperate to eciently disseminate data items corresponding to uploaded

publications towards receivers which have advertised relevant interests.

We will refer to max as the selectivity of the subscription and lifetime can be in-

terpreted as the delay allocated by receivers to the content-based network to satisfy an

interest. The content-based network delivers to receiver r (via home router R), at most

max relevant publications before lifetime expires. Depending on r preferences, content

delivery is either performed in real-time or delayed until lifetime expires at the latest or


until r requests the publications retrieved on its behalf by R. This latter possibility re-

quires to provision a minimum amount of caching resources at home routers, but provides

further temporal decoupling between receivers and their home routers allowing temporary

disconnections of the receivers.

These two variants are depicted by Fig. 2.2. Content delivery from home routers to

receivers require that the former maintain some states regarding their registered customers.

For the sake of clarity, we assume that retrieved publications are delivered to receivers in

real-time.

Receiver Home router

subscribe(predicate, max, lifetime, freshness)

deliver(pub1)

deliver(pubmax)

...

refresh(predicate, max, lifetime, freshness)

lifetime

(a) Real-time content delivery

Receiver Home router

subscribe(predicate, max, lifetime, freshness)

pub1 available

pubmax available

...

refresh(predicate, max, lifetime, freshness)

lifetime

consume

deliver(pub1,...,pubmax)

(b) Delayed content delivery

Figure 2.2: Interactions between receivers and home routers

The service model described above captures the attention span of information consumers

by allowing them to specify the maximum amount of relevant information they would like to

receive over a period of time. Also, it captures the qualitative heterogeneity of information

consumers at global scale.

When freshness equals zero, the interest has the conventional semantic of a subscrip-

tion, i.e. the interest selects only future publications. When freshness and lifetime are

both positive, the interest is a loose subscription which diers from a conventional sub-

scription by the fact that requesters are only interested in publications that they did not

consume previously. Loose subscriptions can be refreshed after lifetime expires. Then,

the system guarantees to the requester that the refreshed subscription is not satised with

previously delivered publications. Finally, in the case where lifetime equals zero, then the

interest is non-persistent, i.e. a request to the mediation network to retrieve immediately

up to max available publications. Table 2.2 provides typical parameter settings for various

applications.

This thesis focuses on the ecient processing of loose subscriptions which have not

been previously studied in the literature. Ecient processing of non-persistent interests

(resp. subscriptions) in an unstructured overlay have been largely studied [YGM02] (resp.

[CRW04] [CRW06]) and previous work can be leveraged. Note that within the framework,

262.2. E-CBN: AN ENHANCED SERVICE MODEL FOR EXTREME SCALE

CONTENT-BASED NETWORKING

Application predicate selectivity lifetime freshness

Notication services (terms="RER+C+Infos") all 7 days 0

News alerting services (type=article, terms="election+2012") 10 24h 24h

Content retrieval (type=article, terms="sophia+perennis") all 24h any

Table 2.1: Examples of interests for dierent applications

non-persistent interests can be processed similarly to loose subscriptions advertised with a

very small lifetime.

Loose subscriptions (freshness > 0 and lifetime > 0)

Non persistent interests (lifetime = 0)

Subscriptions (freshness = 0)

Lifetime

Freshness

Lmax

Fmax

0

Figure 2.3: Illustration of the dierent semantics of the service model

Receivers are allowed to refresh their interests (loose subscriptions) when they expire.

As such, an important requirement for the usability of the service is dened as follows:

Req. 1 The service should guarantee that a refreshed interest will be satised at most once

by any publication over its successive lifetimes. This condition should be enforced without

having to track an exhaustive history of all interests satised with a publication or of all

publications already consumed by a subscription.

To allow routers to dierentiate refreshed interests from new ones, we assume the ex-

istence of an agreement between routers and receivers for such purpose. Control messages

exchanged by routers to advertise interests include a refresh ag indicating whether inter-

ests are refreshed or not (see Fig. 2.7).

Denitions 1 Let S(predicate;max; lifetime; freshness) be a subscription registered at

t0 issued by r, NS be the number of publications notied to r by t0 + lifetime and MS be

the total number of publications uploaded between t0 and t0 + lifetime and matching S.


DPT

OPT

PPT

caching memory

index

forwarding table

PITcache indexes

Boxes table

(Subscribers preferencesand monitoring data)

PIT: Pending interest table

DPT: Disseminated publication table

PPT: Pending publication table

OPT: Overflowing publication table

Buffer

Figure 2.4: Broker model

S is satised when:NS = max: (2.1)

Starvation happens when:NS < max MS (2.2)

Starvation occurs due to congestion or due to the service failing to timely satisfy inter-

ests. The starvation probability, i.e. the frequency of occurrence of starvation, is the metric

used to characterize the quality of service oered by an implementation of the service model

compared to an implementation of the exhaustive ltering semantic. Starvation does not

account for interests which are not satised due to content unavailability.

Problem statement: We aim at minimizing with a low state complexity,the starvation probability and the communication costs required to sat-isfy a workload of loose subscriptions assuming bandwidth, storage andprocessing resource constraints.

2.3 e-CBN: Architecture and mechanisms

In order to implement the service model introduced previously in this chapter, the e-CBN

framework addresses the following design issues :

28 2.3. E-CBN: ARCHITECTURE AND MECHANISMS

PPT

OPTforward garbage

DPT

()

()

()

()

()

()

(): Publication selected by local interest. Enable DF.

(): Publication selected by remote interest and DF is disabled.

(): Publication selected by remote interest.

(): Publication selected by local interests.

(): Publication selected by any interest.

(): Publication selected by a remote interest and DF is enabled.

Figure 2.5: Publication lifecycle inside a broker B

Content forwarding schemes supporting both content retrieval and dissemination traf-

c. Content dissemination trac should be handled by leveraging existing content-

based forwarding schemes. Integrating content retrieval and dissemination in the

same middleware is not straightforward. An important requirement is the ability to

take into account bandwidth and resource constraints by assigning priorities to tran-

siting content. Such forwarding algorithms should minimize starvation in presence of

congestion.

Dissemination strategies dening the conditions making a publication eligible for dis-

semination towards interested receivers. Our service model requires strategies pacing

the dissemination process to information needs.

Interest forwarding strategies necessary to support content forwarding schemes and

dissemination strategies that should take advantage of locality in content availability

patterns as well as attention span quantication in order to minimize communica-

tion costs, unlike existing content-based routing schemes that broadcast subscription

messages to satisfy an exhaustive ltering semantic.

Caching policies increasing content availability as well as communication-eciency.

2.3.1 Router model

The model of a broker is depicted by Fig. 2.4. Each broker needs the following data

structures to operate:


matchforward

P

()

()

()

()

()

PPT: pending publication table

DPT: disseminated publication table

OPT: overflowing publication table

PPT

OPT

DPT

()

()

()

()

()

()

represents exclusive transitions

ending

! : P is a publication requested by B and used to satisfy local interests.

! ! : P is a transiting publication that matches interests advertised by downstream brokers/receivers while the NRT policy is enabled or whileB is a recipient of the publication and PF disabled or P is an uploaded publication matching local and remote interests.

! : P is an uploaded publication matching local interests or no interests.

! ! : P is an uploaded publication selected by remote interests only or P is a transiting publication matching remote interests only, while theNRT policy and PF are enabled.

! : P is a transiting publication with PF enabled matching no interest in the table while the NRT option is enabled or matching already satisedlocal interests.

Figure 2.6: Caching and forwarding decisions inside a broker B

Pending interests table (PIT), which is constituted of an index of interests advertised

to this router and a forwarding table which provides information about the origin of

the interest. The index supports a matching method which provides the identiers of

the interests matching a given publication and the forwarding table associates interest

identiers to the originating broker. We assume that once an interest is satised, the

corresponding home router removes the interest states from the PIT.

Pending publications table (PPT) references pending publications, which have just been

uploaded at some router from publishers and that are waiting for opportunities to be

further disseminated i.e. which have not been used to satisfy remote interests. Pend-

ing publications are referenced in the PPT. A pending publication may have been used

to satisfy interests, which are local to a router. In order to avoid that refreshed in-

terests consume the same publication at their home router, we add a dispatched ag

(DF) in the PPT indicating whether a pending publication has been used to satisfy

local interests or not. Note as well that when a publication is disseminated for the

rst time, it is forwarded with the pending ag (PF) enabled, except when the PDP

option is used (See section 2.3.4.1).


predicate

max

lifetime

freshness

identifier

refresh flag

metadata

identifier

pending flag

payload data

Interest message Publication message

nonce

score

in reply of: interest identifier

The timestamp of an interest message is used to compute the validity of an interest and the timestamp in publication messages is used to compute the age of apublication. Interest messages may include also the origin of the message but this is considered as an implementation detail.

Figure 2.7: Interest and publication messages structure

Overowing publications table (OPT) references overowing publications, which have

been disseminated to remote subscribers, but that never served locally. In fact, they

may be useful for refreshed interests in future lifetimes. New entries are added to

the OPT for pending publications selected by remote interests with dispatched ag

DF disabled and for publications incoming with pending ag PF enabled without

satisfying local interests. The latter situation occurs whenever the content-based

network returns more publications than requested or the states corresponding to

an interest are still in the forwarding table or the en-route caching optimization is

enabled (See Section 2.3.2.2).

Disseminated publications table (DPT) references publications which have been dis-

seminated towards remote routers and used to satisfy local interests.

Boxes table (BT), which references abstractions called box, which are used to keep the

preferences of the consumers as well as to monitor the service oered to them. For

instance, brokers will use boxes to monitor the set of publications selected to satisfy

local interests during their lifetime and use that information to detect duplicates.

Also, they will use the boxes to detect that interests are satised and stop advertising

them in their PIT and optionally advertise overload for that interest in the mediation

network (See Section 2.3.4).

Buers upstream and downstream the forwarding decision modeling the two bottlenecks

of the brokers. The rst bottleneck is related to the matching method of the PIT and

the second bottleneck is related to the transmission of publications (Fig. 2.6).


The cache indexes introduced above have been designed to eectively support Req. 1,

while maximizing opportunities to satisfy refreshed interests. Fig. 2.5 depicts the lifecycle

of a publication inside the cache of a broker i.e. the transitions between the indexes. Fig.

2.6 depicts the processing of an incoming (uploaded or transiting) content by a broker.

Besides the new/refreshed classication, from the perspective of a broker, a remote interest

denotes an interest advertised by downstream brokers, while a local interest denotes an

interest registered by a local receiver.

2.3.2 Caching policies

We assume that brokers are provisioned with a nite amount of caching resources. One

might argue that it is always possible to extend brokers caching resources by distributing

broker behaviors over a cluster of COTS servers as performed today by major online service

providers. In practice, the size of a cluster is limited by practical reasons such as the lack

of space or some energy considerations. Also, the size of the cache impacts the size of

the cache indexes which should ideally t into main memory1. Consequently, beyond some

amount of caching resources, it may be necessary to partition the cache indexes over several

nodes. But, the response time of every broker to satisfy an interest will increase with the

number of nodes involved in the processing and may impact timely delivery of content.

A publication incoming at a brokerB and originating from a receiver or another broker is

cacheable whenever, B is the broker that originally advertised the interest(s) that triggered

the retrieval/dissemination of the content in the mediation network, or B is the broker

where the publication is originally uploaded or, B is allowed to opportunistically cache

transiting publications i.e. when the en-route option introduced in Section 2.3.2.2 is

enabled.

2.3.2.1 Selection and replacement policies

Each router executes a selection policy to determine which publication to select rst to

satisfy an incoming interest, and a replacement policy to determine which publication to

replace rst in case of cache overow. Ideally, selection and replacement policies should

achieve an optimal trade-o between the following tussles: receivers privileging fresh infor-

mation, publishers wanting to reach the widest possible audience with their publications

and network operators willing to minimize communication costs and maximize the quality

of service oered to consumers.

Selection policies should guarantee that new interests are satised with any available

publication, while refreshed interests are served only with publications that have not been

1A common assumption in Information retrieval research


delivered to them during previous lifetimes. In order to be consistent with Req. 1, refreshed

interests should not be satised with publications indexed in DPTs or in OPTs of remote

routers. More precisely, refreshed interests are satised rst with overowing and pend-

ing publications available at the originating router with DF disabled, then with pending

publications available at remote routers of the network.

Consequently, new interests have more opportunities to be satised than refreshed ones.

We can increase content availability for refreshed interests by making pending publications

more persistent. At high-level, selection and replacement policies may discriminate or not

publications according to their type (disseminated, pending, overowing).

We consider two high-level discriminating policies simply called discriminators:

DPF (disseminated publications rst), which returns rst disseminated publications,then overowing and nally pending ones.

PPF (pending publications rst), which returns rst pending publications, then over-

owing and nally pending ones.

We note discriminator-selection policy/discriminator-replacement policy the combination

of policies executed by a broker. The rst discriminator in the notation applies to the

selection policy and the second one applies to the replacement policy. In the case where

no discriminator applies, the discriminator eld of the notation and the following dash are

left blank.

Selection and replacement policies should be designed/chosen in order to balance the

trade-o between new and refreshed interests as well as to meet the expectations of infor-

mation consumers, information providers and network operators:

In order to balance the tradeo between new and refreshed interests, the followingcombinations of policies can be considered: (DPF-*/DPF-*), and (PPF-*/DPF-*) where

* may refer to one of the following policies: most recently used (MRU), least recently

used (LRU), most frequently used (MFU), most fresh (MF) or least fresh (LF). These

policies make pending publications more persistent than other publications, which is

risky as pending publications may correspond to unpopular publications.

Information consumers request most recent publications (freshness) and want theirinterests to be satised (availability). Consequently, selecting most fresh information

rst and replacing least fresh information rst i.e. an MF/LF policy, will make most

fresh publications more persistent in the mediation network. An LF/LF policy is also

worth investigating w.r.t. the availability of fresh information.

Information providers want to reach the widest possible audience. Fairness amonginformation providers in terms of content availability for content of similar popularity


is important in order that some popular content providers do not get discriminated.

An interesting policy to investigate to achieve such fairness is LFU/MFU.

From the perspective of network operators, communication-eciency and contentavailability for consumers can be achieved by enforcing an MRU/LRU policy [JST+09].

In the next chapter, we evaluate the following congurations:

Conguration Expected properties Name

LFU/MFU Fairness LFU

MRU/LRU Communication-eciency and availability MRU

MF/LF Availability of fresh content MF

LF/LF Availability of fresh content LF

DPF-MF/DPF-fLF, LRUg Increases availability of content for refreshed interests DPF, DPFuPPF-MF/DPF-fLF, LRUg Increases availability of content for refreshed interests PPF, PPFu

Table 2.2: Caching congurations to investigate

2.3.2.2 Caching policies

Besides the selection and replacement policies, routers can enforce one of the following

policies:

Default: With this policy enabled, publications are cached only at publishing or requesting

routers.

En-route caching (NRT): With this policy enabled, routers are allowed to cache tran-

siting publications according to the enforced replacement policy and if the content

items are not already present into the cache. The evaluation in the next chapter clar-

ies situations where routers have incentives to cache transiting content. Selecting

publications replicate those publications at several routers and consequently increase

their availability. Replication and persistence of selected publications are expected

to increase when the NRT policy is enabled.

2.3.3 Interest forwarding

Interests have a persistent and/or temporary lifetime. In their temporary lifetime, i.e.

during their propagation in the network, interests are satised with cached publications.

Instead, in their persistent lifetime, interests are satised with pending publications which

have just been uploaded or selected for dissemination. Advertising an interest consists in

ooding the interest in the mediation network until it is satised or all brokers are notied

with the interest.


Note that in order to avoid loops and duplicates with arbitrary topologies, we assume

that interest messages embed a globally unique identier to detect cycles. This is more

convenient than operating over global spanning trees. Designing such scheme is an im-

plementation detail, but it can be achieved by generating hierarchical identiers where

the prex is the identier of the Mediation router. In our work, we assumed that during

their lifetime, interests are advertised once in the mediation network, otherwise a nonce

would have been necessary. Moreover, we assume that interests have dierent identiers

throughout their successive lifetimes.

Routers exchange interests using the following procedure: Upon reception of S(max; )by router R, if the interest identier has already been advertised in the PIT, then drops the

interest. Otherwise, if the number of relevant publications available into the cache exceeds

or equals max, then max publications are selected for delivery and the propagation is

stopped. Otherwise, S is further advertised with the max parameter decremented by the

number of matching publications oered by R (See g. 2.8).

Fig. 2.8 illustrates a scenario where starvation may occur: Three publications relevant

to I are available but only two of them are forwarded to BrokerC . This is due to the bad

selection decision of BrokerP which returns c4 instead of c5 that would have contributed

to satisfy the interest. This situation underlines the sensitivity of selection policies on

starvation. Moreover, forwarding c4 in the mediation network generates an overhead that

should not be neglected. Note that the requesting broker can always detect duplicate

publications occuring during a lifetime retrieved from the network for the same interest

using the states available in the BT.

In order to attenuate that overhead, we introduce in Section 2.3.6.2 the in-network

duplicate dropping (IDD) mechanism, where a broker drops a publication already present

in the cache because it may have already responded with that content. In order to reduce

starvation due to bad selection decisions, we propose in Section 2.3.6.2, the proactive dupli-

cate dropping (PDD) mechanism or duplicate avoidance, where interests are forwarded with

the list of publication identiers already used to satisfy them along the forwarding path.

These schemes are not perfect. They may infer themselves starvation and an overhead of

publication messages in scenarios such as the one described by Fig. 2.10. Note that both

schemes can be combined.

2.3.4 Dissemination strategies

We discuss below, two simple yet eective strategies in the trade-o between satisfaction

and communication-eciency.


BrokerC BrokerPConsumer

subscribe(I(ps, 3, T, f))

deliver(c1, c4)

forward(c4)

advertise(I(ps, 1, T, f))

{c1, c4} {c4, c5}

cache c4 if possible, otherwise drop it

T

select c4

starvation

{c2, c3}cache replacement

Note: We assume that c1, c4, c5 are relevant to interest I. BrokerC is the

home router of Consumer and the cache replacement event may correspond to

the upload of new content forcing the cache replacement policy to apply.

overhead

Figure 2.8: Starvation illustration

Push/pull and explicit overload notication (EON)

Overload corresponds to the situation where the number of publications retrieved from

the network w.r.t. an interest exceeds the selectivity of the subscription. A publication

P is disseminated by a router R, whenever at upload time, P matches an interest in the

forwarding table or if a matching interest transits through the node while the publication

is pending (push/pull). When an interest is satised i.e. max or more publications have

been retrieved, the corresponding home router advertises an overload notication message

in order that remote routers remove the corresponding states from their tables (overload

notication). Fig. 2.9 describes the interest forwarding strategy operation with the EON

mechanism enabled.

2.3.4.1 Pull/delayed push (PDP)

Unlike EON, PDP does not use explicit notication messages to notify remote routers of

overload. PDP uses the propagation of new and refreshed interests by the content-based

routing protocol to pace the dissemination process (pull) instead of pushing publications

additionally. In order to avoid that starvation occurs in some cases, we compute a most

lately publication time for each uploaded publication matching pending interests. Let P be

a publication uploaded at a router R at time t0 and SP be the set of matching interests

advertised in R's forwarding table (excluding local interests), the most lately publication

time tP associated to P is given by the relation:


BrokerC BrokerPConsumer Provider


deliver(c1)

upload(c2)

{c2}

{c1}

forward(c2)

deliver(c2)

advertise overload(I)

refresh(I(ps, 2, T, f ))


{}

Note: We assume that c1 and c2 are relevant to I.

T

No relevant content available

Remove states corresponding to I

Figure 2.9: Interest forwarding and content dissemination

tP = t0 + minS2SP

(deadline(S)) : (2.3)

being a system parameter and deadline(S) being the remaining time before S expires.

At last, any publication that may contribute to satisfy an interest not yet satised, is

nally disseminated. Note that should be engineered such that it is greater or equal to

the time required to forward a publication between two endpoints of the network. In the

case where a scheduled publication is eligible for replacement by another one, the scheduled

publication is immediately disseminated and the incoming publication cached.

A publication scheduled with PDP is disseminated at scheduling time even though a new

interest selects the publication before its dissemination. Once scheduled, the publication

stops being pending, but overowing or dispatched, depending on whether there exists

local/remote recipients. Thus, while the publication is scheduled it may be selected by

other interests and forwarded into the mediation network. Consequently, publications

disseminated with PDP are forwarded with the pending ag disabled.

2.3.5 Publication forwarding

Publications are forwarded in the content-based network depending on whether they are

selected by pending or transiting interests. In Fig. 2.7, the nonce eld is used to iden-

tify publication messages and the identier eld is used to identify content embedded in

publication messages. Nonces are used to detect duplicates on cyclic topologies. We use

nonces embedded in publication messages in order to avoid forwarding loops instead of

global spanning trees.


BrokerC BrokerPConsumer


forward(c1, c2)


{c1, c2}

T

c2 replaced by c3

BrokerI

{c2}{}

deliver(c1, c2)

forward(c2) {c1,c3}

forward(c2)drop c2

{c1, c2}

overhead


Note: We assume that only c1 and c2 are relevant to I.

Figure 2.10: Duplicate dropping mechanism limitations

Unicast delivery: A publication selected by a transiting interest is forwarded on the

reverse path only towards the requesting broker. The in reply of eld of the cor-

responding publication message (see Fig. 2.7) is initialized to the identier of the

interest that selected the publication and triggered the forwarding of the publication

message. The information of the broker that advertised a particular interest identier

is provided by the PIT.

Multicast delivery (content dissemination): When a pending publication is selected

by pending interests, it is disseminated towards all the receivers that have advertised

the interest. Each broker which belongs to the dissemination tree DT of the pending

publication rooted at the publishing broker, can determine the next hop in DT by

evaluating the publication description against the set of interests advertised in the

PIT. Content dissemination can be realized according to baseline content-based for-

warding (section 2.1), either by performing matching on a hop-by-hop basis or by

performing matching at the publishing broker and switching on the subsequent nodes

of the delivery tree such as the DV/DRP protocol [HCW+06] described in section 2.1.

The latter case would require to encode the in reply of eld of publication messages

(Fig. 2.7) as a type-length-value (TLV) element.


2.3.6 Forwarding optimizations

2.3.6.1 Congestion-aware forwarding

Congestion may appear at some routers. It is important in these conditions to maximize

the satisfaction of the service. This can be done by assigning priorities to transiting publica-

tions. The priorities are used to schedule publications in buers upstream and downstream

the forwarding decision.

For this purpose, we compute a score for each publication embedded in publication

messages. The score is recomputed by every hop on the delivery path. Let R be a router

and be the set of publications buered at R that are waiting to be further disseminated.

For each publication P 2 , we dene SP the set of interests matching P downstream R,and compute a score Cbf(P) giving higher priority to popular and urgent publications.

We estimate popularity by the number of matching interests advertised downstream R and

urgency by the minimum deadline among matching interests. Cbf(P) is dened by the

following equation:

Cbf(P ) =jSpj

minI2SP

(deadline(I)): (2.4)

Note that in the unicast case, the score associated to a publication selected by an

interest I is simply 1=deadline(I). Obviously, in any case, to avoid a division by zero in

equation 2.4, interests whose deadlines have been reached are removed from the PIT and

not considered in the computation of the score.

2.3.6.2 Handling duplicate responses

When an interest is advertised in the content-based network, dierent brokers may reply

with the same (overowing or dispatched) publication. In the worst-case, the duplicate

responses will be detected by the requesting broker. This may infer starvation in scenarios

such as the one depicted by Fig. 2.8, as well as a communication overhead. In order, to

mitigate that overhead, we propose the mechanisms described below.

In-network duplicate dropping: Every broker, upon the arrival of a publication,

checks whether the publication already appears in its cache and if this is true, it dis-

cards it. Otherwise, it forwards the publication according to the technique described in

Section 2.3.5. The reason for searching the cache of a broker upon the arrival of a pub-

lication, is because publications follow the reverse path that interests follow. This means

that the interest that selected the corresponding publication has also been processed by

the broker under question which may have responded to that interest with the same cached

publication(s). Note that this mechanism is not always eective as depicted by Fig. 2.10:


The intermediate router BrokerP cannot detect that content c2 has already been forwarded

to BrokerC (that itself forwarded). As a consequence, BrokerP redundantly forwards c2 to

BrokerC which nally drops it thanks to the states installed in the BT.

Note that the duplicate dropping heuristic does not thoroughly solve the problem ad-

dressed but are expected to alleviate the communication overhead. Fig. 2.10 exhibits

extreme cases where in-network duplicate generates starvation and communication over-

head.

Duplicate avoidance: In order to improve the eectiveness of the in-network duplicate

dropping scheme, we propose a proactive counterpart such that a broker processing an

interest and selecting cached publications to serve an interest, append the list of identiers

of the selected publications to the corresponding eld in the interest message before further

forwarding the interest. With this mechanism, duplicate responses will be eliminated from

every branch of the tree within which the interest is forwarded.

2.4 Discussion

This chapter described the E-CBN framework for extreme-scale content-based networking.

The framework introduces a service model capturing the quantitative and qualitative het-

erogeneity of information consumers and addresses the key design issues in implementing

eciently the service model. In the next chapter, we quantify the communication gains

of E-CBN over baseline content-based networking under realistic workload assumptions and

evaluate the eectiveness of each mechanism implemented by the framework.

40 2.4. DISCUSSION

Chapter3Evaluation

Contents

3.1 Workload characterization . . . . . . . . . . . . . . . . . . . . . . 42

3.1.1 On the relevance of workload modelling . . . . . . . . . . . . . . . 42

3.1.2 Evaluation methodology . . . . . . . . . . . . . . . . . . . . . . . . 43

3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

3.3 Caching policies evaluation . . . . . . . . . . . . . . . . . . . . . . 50

3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

41

42 3.1. WORKLOAD CHARACTERIZATION

This chapter provides a characterization of the framework with respect to the quality of

service oered to consumers and communication-eciency. Firstly, we quantify the gains

of the framework (e-CBN) over a baseline scheme implementing the exhaustive ltering se-

mantic (EF). We assume in the EF case that receivers caching resources are unlimited which

is an extremely unfavorable assumption for e-CBN. Secondly, we evaluate the eciency of

the eight policies introduced by table 2.3.2.1 and discuss the benets of en-route caching

(NRT). Finally, we measure the performances of the congestion-aware forwarding scheme.

Note that in the experiments, we did not observe any occurrence of duplicates ltered at

the edge. Consequently, it is worthless to evaluate the IDD and PDD schemes.

Performances are characterized using the following metrics:

Satisfaction, which is the percentage of interests satised as dened by Eq. 2.1. Weuse this metric mainly to compare an instance of the framework to EF.

Starvation probability (SP), which is the frequency of occurrence of starvation. SPis preferred when comparing two instances of the framework. Unlike the semantics of

satisfaction as dened above, SP does not account for interests which are not satised

due to the unavailability of relevant publications. SP accounts only for interests which

are not satised, while EF would have satised the interest assuming unlimited caching

resources at the receivers.

Message trac, which is the total number of publication messages forwarded intothe mediation network. We also characterize the message trac by the bandwidth

saved, which is how much bandwidth has been saved for a given scenario using e-CBN

instead of EF.

Control trac, which is the number of interest messages forwarded into the me-diation network. We characterize the control trac also by the control overhead,

which is the ratio between the control trac generated by an instance of e-CBN and

the control trac generated by an instance of EF.

3.1 Workload characterization

3.1.1 On the relevance of workload modelling

When evaluating content-based publish/subscribe approaches, workload assumptions in

terms of popularity and locality signicantly impact measured performances. Due to the

lack of publicly available datasets of large scale content-based publish/subscribe systems,

synthetic workload generation has been widely accepted. The resulting challenge is how to

CHAPTER 3. EVALUATION 43

choose the parameters of the workload model in order to generate a workload consistent

with a given application prole.

For instance, content-based routing (CBR) [CRW04] is preferable over pure broadcast

only in scenarios with a sparse density of receivers. For this reason, previous CBR research

has been evaluated under specic popularity and locality assumptions.

The seminal work on content-based routing [CRW04] is evaluated assuming that the

density of receivers equals 75%, and the popularity distribution is characterized by the

matching message distribution, which represents the number of messages matching a per-

centage of predicates. Most messages match 5% to 15% of the predicates, a signicant num-

ber of messages do not match any predicate, and no message matches more than 25% of the

predicates. Another signicant work, Kyra [CS04] was characterized for comprehensiveness

with four dierent popularity and volume distributions reported in the publish/subscribe

literature.

Majumder et al. study in [MSRS09] the impact of locality patterns on the communication-

eciency of several clustering algorithms for content-based routing. The workload is tuned

from a localized subscription model, i.e. similar subscriptions originating from the same

region, to an uniform model. In fact, the locality of similar subscriptions has a signicant

impact on the eciency of multicast-based schemes and the eciency of optimizations such

as subscription covering [TK06].

The sensitivity of topologies on the communication-eciency of content-based routing

schemes is discussed in [MC07], where a run-time algorithm to adapt the topology to the

application demand is proposed.

A very interesting fact about the importance of workload assumptions on the conclusions

one can draw from the evaluation of its solution is the positioning of Riabov et al. in

[RLW+02] regarding the conclusions on the benets of multicast in one of the Gryphon

papers [OAA+00]. Riabov et al., demonstrated that the conclusions from the Gryphon

papers were not always true and that they depend on the assumptions made on locality

and similarity properties.

3.1.2 Evaluation methodology

For the evaluation of the framework, we assume a news dissemination application such

as Google alerts implemented with e-CBN distributed over a network of brokers. The

workload is generated in order to meet this application prole. In order to validate our

algorithms and design choices, we implemented the framework in the PEERSIM simulator

which is a common choice for the evaluation of large-scale publish/subscribe solutions.

Up to 8 GB of RAM memory were necessary to run the experiments given the size of the

forwarding and caching indexes.


We consider that a set of events generate both publications and interests, and that three

parameters characterize each event: popularity, locality and volume. The popularity of an

event refers to the number of interests related to it, its volume to the number of related

publications and its locality to the regions of the topology likely to originate related interests

and publications. A similar methodology is used in [CS04]. We prefer this methodology

because it ts the target application model, and because it allows an easier control of

popularity and locality assumptions.

In the generation of our workload, we assume an arbitrary router topology of 100 nodes

characterized by average degree 4, diameter 6 and following a power-law distribution with

few nodes of high degree and many nodes of small degree (See Fig. 3.1). We assume that

each broker is provisioned with some amount of caching resources and can cache up to CS

objects.

Figure 3.1: Evaluation topology

We assume that most popular events are characterized by larger and broader audiences in

terms of number of interests, and are also likely to trigger larger volume of publications. In

other terms, event popularity and volume are generated according to the same distribution.

Remember that we evaluate only loose subscriptions.

We investigate two dierent locality patterns for the workload generation:

Random: Publications and interests originate from random locations.

Realistic: Most popular events are more widely spread in the topology.

Let us consider ei (1 i E), an event of popularity pi, volume vi and locality li.pi is sampled from a power-law distribution and the volume vi is such that vi = P pi,where P is the total number of publications generated by the scenario. When the random


Date Monitored events Volume Peak popularity

12-04-2011 33 9507 30%

13-04-2011 31 7941 40%

14-04-2011 29 8875 45%

Table 3.1: Statistics from the front page of Google news

Parameter Default Value

Dissemination policy EON

Number of interests 100 000

Number of publications 50 000

Cache size (CS) 5000

Refresh probability 1.0

Number of events 50

Zipf exponent 1.0

Buer size 50000

Average selectivity 10

Simulation time 100

Average lifetime 20

Freshness 50

Network Size 100

Locality pattern realistic

Selection/replacement policy MF

Caching policy Default

1

Table 3.2: Default parameter values for the evaluation

locality pattern is enforced, publications and interests associated to ei can be issued by any

of the N routers constituting the mediation network. When the realistic locality pattern

is enforced, publications and interests associated to ei can be issued only from a set of

nodes computed using li. In that case, we dene li such that li = pi and such that dli Nerouters are potential issuers (hosting interested receivers) of interests related to ei. This set

of routers is computed by choosing a random root node and dli Ne 1 additional nodesamong the closest routers to the root. We assume a constant arrival rate for interests

(resp. publications) equals to rs = S=T (resp. rp = P=T ), where S is the total number of

subscriptions generated by the scenario.

For each publication (resp. interest), we randomly select an event and a location among

the routers eligible to generate trac related to that event. When the volume (resp.

popularity) associated to an event is reached, it is removed from the set of events that can

be used to generate new publications (resp. interests). We generate selectivity and lifetime

values associated to interests according to Poisson distributions of average reported by

table 3.2. Randomizing selectivity and lifetime values is necessary to model heterogeneous


consumer requirements.

We assume that each interest is refreshed at the end of its lifetime with a probability Pr

called refreshed probability, otherwise a new interest related to the same event is generated.

For the simulation of PDP, we set to 1, which is an upper bound of the end-to-end transfer

delay between any pair of brokers in absence of bottleneck.

Simulation time is set to 100. Consequently, every time unit, 500 new publications

are uploaded followed by the generation of 1000 new interests. Note that the PEERSIM

simulator does not provide any guarantee about the order in which the scheduler processes

the scheduled events.

Evaluation settings In order to model the event popularity distribution used to generate

the trac, we measured the volume distribution of the top stories reported by the french

front page of Google news during three consecutive days. It is interesting to note that on

the 14th April, the top story reached a popularity of 45% (see table 3.2) illustrating the

fact a single event may represent a major fraction of trac.

5 10 15 20 25 300

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

event ID

volu

me

120420111304201114042011

(a) Volume distribution

5 10 15 20 25 300

0.05

0.1

0.15

0.2

0.25

0.3

event ID

12042011zipf(1.0)zipf(0.7)zipf(2.4)

(b) Volume distribution tting (I)

5 10 15 20 25 300

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

event ID

volu

me

14042011zipf(1.0)zipf(0.7)zipf(2.4)

(c) Volume distribution tting(II)

Figure 3.2: Measures from Google news

Fig. 3.2 shows that the measured volume distribution can be approximated by a power-

law distribution of exponent 1.0 (12th april) or 2.4 (14th april). For workload generation, we

choosed exponent 1.0 in order to avoid a too much favorable scenario, where a single event

would generate most of the trac. The peak popularity equals 30% followed by a peak

of 10% i.e. the two most popular events generate 40% of publications and subscriptions

trac. We use the same series of event popularity in all our experiments. In fact, Fig.

3.3(c) shows that there is much volatility in the generated series and demonstrates the

importance to use the same series for the comprehensiveness of the results.

Regarding the refresh probability, we choose Pr = 1:0 i.e. each interest is automatically

refreshed after expiration. This is inline with the fact that RSS trac has been reported to

be sticky contrarily to web trac [LS05]. The number of events and publications simulated

has been sized on the basis of measures reported in table 3.1. Without explicit mention,

the reader should refer to table 3.2 for the default parameters values.


Workload characterization We characterize the workload using three distributions:

matching distribution, density of recipients and gap distribution. The matching distribution

represents the CCDF of the popularity distribution of publications, which associates each

publication to the percentage of matching interests. The density of recipients represents

the CCDF of the topological popularity distribution of publications, which associates each

publication to the percentage of brokers which have requested the publication. The gap

distribution characterizes the dierence existing between consumers' attention span and

the volume of information available.

0 0.005 0.01 0.015 0.02 0.0250

0.2

0.4

0.6

0.8

1

popularity

proba

bili

ty

(a) Matching distribution

0 0.2 0.4 0.6 0.8 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

density of recipients

proba

bili

ty

random localityrealistic locality

(b) Density of recipients

0 10 20 30 40 500

0.05

0.1

0.15

0.2

0.25

0.3

0.35

event ID

(c) Popularity distribution

0 2000 4000 6000 8000 100000

0.05

0.1

0.15

0.2

0.25

0.3

0.35

gap values

(d) Gap distribution

Figure 3.3: Workload characterization

Fig. 3.3(b) characterizes the density of recipients and shows that no publication is

requested by more than 24 % of brokers with realistic locality enabled. Conversely, with

random locality enabled, most publications have a topological popularity superior or equal

to 90 %. Realistic locality provides a less extreme scenario than random locality. Moreover,

this is inline with what has been reported from [LFHKM05], that for 60 % of les in an

e-Donkey network, more than 80 % of the replicas were in the same country. Fig. 3.3(a)

shows that in the workload there is no publication with popularity larger than 2.5%. The

gap distribution depicted by Fig. 3.3(d) is positive and clearly describes a situation of

information overload.

3.2 Results

Comparison of an instance of the framework to a variant of EF First, we compare

an instance of e-CBN to EF for dierent load levels obtained by varying the selectivity and

with realistic locality enabled.

48 3.2. RESULTS

In Fig. 3.4(a), e-CBN satises 13% more interests than EF for various load levels.

e-CBN satises more interests than EF because e-CBN use cached publications to satisfy the

interests while EF doesn't take advantage of the cache. Consequently, more publications

are available with e-CBN than with EF. Fig. 3.4(b) shows that for small selectivity values,

e-CBN saves almost 100% of bandwidth and up to 40% of bandwidth for larger selectivity

values. These gains are obtained while still generating less control trac than EF.

Fig. 3.5 shows that with the MF policy and larger cache sizes, more interests are satised

(Fig. 3.5(a)), and communication-eciency is improved. Also, there is no incentive to add

caching resources beyond 1% of the total volume of publications. This is not a surprising

result, since with more caching resources, brokers have more opportunities to satisfy local

interests with cached publications. Note that Fig. 3.5(a) does not display the percentage

of satised interests by EF, because it is obvious that EF does not take advantage of the

cache to satisfy more interests.

Finally, 3.6(a) shows that the when the size of the network increases also increases the

satisfaction ratio of the two methods. Of course this increase is not very signicant (double

satisfaction ratio when the number of nodes is six times larger). More nodes/brokers might

mean more dierent cached messages, but on the other hand more brokers also increase

the replication degree of the cached messages which does not allow the retrieval of unique

cached items. At any case we observe that e-CBN performs on average 10%-34% better

than the EF. Also, it is obvious that for the default selectivity value the e-CBN saves almost

88% of bandwidth for small networks and up to 98% of bandwidth for larger network (see

Figure 3.6(b)), while generating less control trac (see Figure 3.6(c)).

0 %

20 %

40 %

60 %

80 %

100 %

5 10 15 20 25 30 35 40 45 50

satisfaction

selectivity

e-CBNEF

(a) Satisfaction

0 %

20 %

40 %

60 %

80 %

100 %

5 10 15 20 25 30 35 40 45 50

bandwidth saved

selectivity

(b) Message trac

0 %

20 %

40 %

60 %

80 %

100 %

5 10 15 20 25 30 35 40 45 50

control overhead

selectivity

(c) Control trac

Figure 3.4: Performance comparison of an instance of e-CBN to EF for dierent load levels

Dissemination methods evaluation Fig. 3.7 compares the performances of PDP to

EON for dierent load levels. We observe that PDP and EON display close performances

with respect to the quality of service (less than 2% of starvation even for extremely large

selectivity values) and the amount of trac generated by the content-based network. It is

noticeable that PDP generates slightly less control trac than EON. This may be due to the


20 %

25 %

30 %

35 %

40 %

0 100 200 300 400 500 600 700 800 900 1000

satisfaction

cache size

(a) Satisfaction

60 %

65 %

70 %

75 %

80 %

85 %

90 %

95 %

100 %

0 100 200 300 400 500 600 700 800 900 1000

bandwidth saved

cache size

(b) Message trac

88 %

90 %

92 %

94 %

96 %

98 %

100 %

102 %

104 %

106 %

108 %

110 %

0 100 200 300 400 500 600 700 800 900 1000

control overhead

cache size

(c) Control trac

Figure 3.5: Performance comparison of an instance of e-CBN to EF for dierent cache sizes

25 50 75 100 125 150 175 200 225 250 275 300

0%

10%

20%

30%

40%

50%

60%

70%

80%

e-CBN

EF

satisfaction

network size

(a) Satisfaction

25 50 75 100 125 150 175 200 225 250 275 300

86%

88%

90%

92%

94%

96%

98%

100%

bandwidth saved

network size

(b) Message trac

25 50 75 100 125 150 175 200 225 250 275 300

75%

80%

85%

90%

95%

100%

105%

110%

control overhead

network size

(c) Control trac

Figure 3.6: Performance comparison of an instance of e-CBN to EF for dierent networksizes

fact that PDP does not generate overload notication messages.

Enhanced forwarding scheme evaluation We compare the performances of the en-

hanced forwarding scheme (CBF) to a simple FIFO policy for dierent buer sizes. Fig.

3.8(a) shows that FIFO achieves a bad quality of service and does not take advantage of

larger buer sizes. CBF instead takes better advantage of the available bandwidth and

processing capacity to improve the QoS oered to consumers.

50 3.3. CACHING POLICIES EVALUATION

0.001 %

0.01 %

0.1 %

1 %

10 %

100 %

5 10 15 20 25 30 35 40 45 50

starvation probability

selectivity

EONPDP

(a) Starvation

0

0.5

1

1.5

2

2.5

3

3.5

5 10 15 20 25 30 35 40 45 50

message traffic (million of messages)

selectivity

EONPDP

(b) Message trac

0

5

10

15

20

25

30

5 10 15 20 25 30 35 40 45 50

control traffic (million of messages)

selectivity

EONPDP

(c) Control trac

Figure 3.7: Performance comparison of EON and PDP

0 %

5 %

10 %

15 %

20 %

25 %

30 %

100 200 300 400 500 600 700 800 900 1000

starvation probability

buffer size

FIFOCBF

(a) Starvation

0.1

1

10

100

1000

100 200 300 400 500 600 700 800 900 1000

message traffic (thousand

Documents

These Diallo