Parallel Multi-path Forwarding Strategy

for Named Data Networking

Abdelkader Bouacherine, Mustapha Reda Senouci and Billal Merabti

A.I. Laboratory, Ecole Militaire Polytechnique,

P.O. Box 17, Bordj-El-Bahri 16111, Algiers, Algeria

Keywords:

Named Data Networking, NDN Flow, NDN Fairness, Forwarding Strategy, Forwarding Decisions, Weighted

Alpha-proportional Fairness, Flow Assignment Problem.

Abstract:

Named Data Networking (NDN) is one of the most promising instantiations of the Information Centric Net-

working (ICN) philosophy. This new design needs a new thinking due to the fact that the deﬁnitions of some

concepts used in TCP/IP paradigm are no longer appropriate. In this context, ﬂow and fairness concepts are

examined and new perspectives are proposed. An important literature exists about forwarding strategies and

congestion control in NDN context. Unfortunately, the lack of deﬁnitions pushed many researchers to use the

TCP/IP heritage. As a consequence, they neither fully beneﬁt from the native multi-path support nor address

the fairness problem. In order to overcome such a drawback and to meet end-users fairness while optimizing

network throughput, a new Parallel Multi-Path Forwarding Strategy (PMP-FS) is proposed in this paper. The

PMP-FS proactively splits trafﬁc by determining how the multiple routes will be used. It takes into consid-

eration NDN in-network caching and NDN Interest aggregation features to achieve weighted alpha fairness

among different ﬂows. Obtained preliminary results show that PMP-FS looks promising.

1 INTRODUCTION

Internet is built on redundancy in both communica-

tion and information networks. Exploiting this multi-

plicity is a necessity and an obligation nowadays and

in the future Internet (Qadir et al., 2015). With the

Information Centric Networking (ICN) proposals, the

inherited constraints imposing the single-path routing

that limit the utilization of Internet network multi-

plicity have disappeared. Named Data Networking

(NDN) (Jacobson et al., 2009; Zhang et al., 2010;

Afanasyev et al., 2014) is a future Internet architec-

ture proposal rolling under the ICN paradigm. The

NDN comes with a new communication model based

on four main characteristics:

1. Receiver-driven data retrieval model: The user

expresses an Interest with a uniquely identiﬁed

name. The routers use this latter to retrieve the

data whose name matches the requested one, and

return it to the user;

2. Local state information decisions: They are based

on the kept local state information. No global

knowledge exists;

3. Loop free forwarding plane: The Interest carries a

nonce and the corresponding returned data packet

follows the same path in the reverse direction.

Thus, none the Interests and data objects can loop

for a sufﬁcient period of time;

4. One-to-one ﬂow balance: One Interest brings

back at most one data object.

The above communication model led to an adap-

tive forwarding plane. The latter combined with an-

other innovative feature (i.e. the NDN in-network

caching) constitute a platform for multi-path support

called ”NDN native multi-path support”. This fea-

ture can be used to take advantage of the Internet

multiplicity in a parallel (simultaneously) or in a se-

quence manner (serial) as a backup conﬁguration (de-

tect failure and retry alternative paths). NDN native

multi-path support is used to optimize the efﬁciency

(throughput) on one hand, and ensures that end-users

get a fair share (fairness) on the other hand.

Serving end-users in a fair manner while maxi-

mizing network throughput is fundamental and a chal-

lenging task in designing a forwarding plane. The

TCP/IP paradigm was built on end-to-end communi-

cation model. In this model, the source and the desti-

Abdelkader, B., Senouci, M. and Merabti, B.

Parallel Multi-path Forwarding Strategy for Named Data Networking.

DOI: 10.5220/0005964600360046

In Proceedings of the 13th International Joint Conference on e-Business and Telecommunications (ICETE 2016) - Volume 1: DCNET, pages 36-46

ISBN: 978-989-758-196-0

nation are known. Packets in the network are tagged

by this information. As a consequence, the fairness

between end-users is directly related to the fairness

between the ﬂows. With this legacy, we inherited

most of the deﬁnitions that we use in nowadays com-

munications and networking community.

Identifying the end-users packets in NDN (In-

terests and data objects) is meant to be impossible

in-between within the network. This is the main

difference between NDN and the TCP/IP paradigm

caused by the disappearance of end-to-end commu-

nication model. The NDN architecture is location-

independent. Therefore, the NDN Interests can be

forwarded individually and independently and data

can be duplicated anywhere and so many times in the

network (NDN in-network caching feature).

The location-independent nature of NDN, the in-

network caching, and the Interest aggregation features

brought by NDN are a shifting innovation in network-

ing. This made the rich literature about multi-path

forwarding strategies, fairness and congestion control

for the TCP/IP architecture no longer appropriate. Al-

though, an important literature exists in NDN context,

the lack of deﬁnitions forced many researchers to use

the TCP/IP heritage. The unpredicted Interest aggre-

gation within the network and the important variation

in RT T (Round Trip Time) measurements, as a result

of the in-network caching feature, prevent achieving

adequate fairness and handling the dynamics in re-

turning data (Yi, 2014). Thus, the native multi-path

support has not been taking advantage of.

To design a multi-path forwarding strategy, one

has to overcome two barriers. First, it is a necessity to

deﬁne the NDN ﬂow independently of the source and

destination of Interests and data. Second, it is manda-

tory to maintain end-users fairness and to adopt ap-

proaches without relying on RT T measurements.

In this context and in order to beneﬁt of NDN na-

tive multi-path support, in this paper, we make the

following contributions:

• We explain and clarify ambiguities of the inher-

ited deﬁnitions of ﬂow and fairness and we pro-

pose new deﬁnitions (perspectives) in the context

of NDN (section 3).

• We formulate the parallel multi-path forwarding

packets as an optimization problem for Coordi-

nated Multi-Path Flow Control (section 4).

• We propose a new localized Parallel Multi-Path

Forwarding Strategy (PMP-FS) (section 4). The

PMP-FS proactively splits trafﬁc by determining

how the multiple routes will be used. It takes into

consideration the NDN in-network caching and

the NDN Interest aggregation features to achieve

weighted alpha fairness among different ﬂows

(Mo and Walrand, 2000).

• We report and discuss PMP-FS preliminary re-

sults (section 5).

2 BACKGROUND AND RELATED

WORK

Internet is built on redundancy. It is a system built on

shared resources. As a system, it seeks to optimize its

efﬁciency (throughput) on one hand, and because of

its sharing nature, it seeks to ensure that end-users get

a fair share (fairness), on the other hand.

The native support of parallel multi-path forward-

ing in NDN is to be taken advantage of. Indeed, the

NDN Forwarding Information Base (FIB) entry con-

tains a list of ranked next-hops for a speciﬁc name

preﬁx to a fetched data object with a unique name.

This data object can be duplicated (cached) anywhere

and so many times in the network. These characteris-

tics lead to a competition between outgoing Interests

as to be forwarded to the set of best possible face(s)

while preserving fairness between end-users.

2.1 Background

Fairness is a crucial concept in networking that needs

complementary information to be understood, since it

varies from equality to equity and can be deﬁned in

many ways. It has therefore already been the subject

of intensive research. Many deﬁnitions were given

such as Weighted Proportional Fairness (Kelly, 1997),

Proportional Fairness (Mazumdar et al., 1991), and

the Max-Min Fairness (Hahne, 1991). A unifying

mathematical model to fair throughput was ﬁrst in-

troduced in (Mo and Walrand, 2000). The so-called

alpha-fair utility functions U

) (Equation 1), which

deﬁnes the different notions of fairness depending on

the choice of a parameter α . This latter is a kind of

degree of fairness that controls the trade-off between

efﬁciency and fairness. Please refer to Table 1 for no-

tations.

) =







1−α

, i f α 6= 1

log(x

) , i f α = 1

(1)

f or w > 0, α ≥ 0, l ∈ L

For example, α = 0 with w = 1 corresponds to the

system maximum efﬁciency or throughput, α → ∞

with w = 1 corresponds to the Max-Min fairness, and

α = 2 with w =

RT T

corresponds to the TCP fair

(Low, 2003).

Parallel Multi-path Forwarding Strategy for Named Data Networking

Table 1: Notations.

Notation Meaning

L set of resources (arcs)

P set of paths or routes

l ∈ P resource l is on path p

capacity of resource l

AX ≤ C capacity constraints

S set of ﬂows

s individual ﬂow (session)

weight of ﬂow (session) s

rate of ﬂow (session) s

ﬂow rate on path p

) alpha-fair utility functions

q a number close but less than 1

α a parameter reﬂecting the fairness

f l number of active ﬂows

shadow price, or rate of

congestion indication at resource l

The proposed fairness schemes used in the TCP/IP

paradigm with enough bandwidth demand allocates

equal bandwidth to the active ﬂows. As a conse-

quence, the fairness between end-users is directly re-

lated to the fairness between the ﬂows. It has worked

because we had a sharp deﬁnition of the ﬂow. It is not

the case of NDN.

2.2 Related Work

NDN differs mainly from the IP paradigm by the dis-

appearance of the end-to-end communication model.

NDN is source-destination free, where NDN routers

have no idea of the provenance of Interests and their

possible destinations. Speaking of end-users fairness

passes certainly through a clear deﬁnition of the NDN

ﬂow.

In (Wang, 2013), the author gives a deﬁnition

of NDN ﬂow as triplet [( group of Interests, data

fetched by this group), client, name-preﬁx]. He sup-

poses that the name-preﬁx is indivisible in the rout-

ing/forwarding domain and the Interests are not for-

warded individually. The author claims that the difﬁ-

culty of deﬁning a ﬂow is not due to the multi-homing

and NDN in-caching feature. We do not agree with

the client component, since NDN routers ignore com-

pletely the sources and destinations of the trafﬁc. We

claim as the authors of (Yi et al., 2012; Yi et al.,

2013; Yi, 2014) did, that fairness in NDN context can

be only deﬁned by the names preﬁxes carried by the

packets.

Let us take the deﬁnition of a ﬂow presented in

(Wang, 2013) and the max-min fairness presented in

(Yi, 2014). At the equilibrium state, the max-min

fairness allocates equal bandwidth to the active ﬂows.

Let us have the example of ”www/com/Facebook/”

as a name preﬁx or as a delegation name (Afanasyev

et al., 2015a), representing one FIB entry under which

we have 9999 pending Interests. The latter name

preﬁx will get the same bandwidth share with an-

other name preﬁx of 99 pending Interests. Most re-

searchers might consider, as we do, this equality as

unfair. Deﬁning the NDN ﬂow to maintain end-users

fairness is mandatory. In the next section, we present

deﬁnitions of ﬂow and fairness in the context of NDN.

3 DEFINITIONS

To deﬁne an NDN ﬂow two questions should be an-

swered:

1. How NDN routers treat names?

2. How NDN routers treat different types of data?

To answer the ﬁrst question, let us examine the

NDN Forwarding Information Base (NDN FIB) func-

tionalities. The NDN FIB entry contains a list of

ranked next-hop for speciﬁc name preﬁx. Due to scal-

ability issues (Baid et al., 2012; Narayanan and Oran,

2015), consensus has not yet been reached on how

and what preﬁx names to populate the FIB. They are a

subject of intensive research (Afanasyev et al., 2015b;

Afanasyev and Wang, 2015; Song et al., 2015; Yuan

et al., 2012). All mentioned proposals advocate that

NDN packet is forwarded by performing a look-up

of its content name using the Longest Preﬁx Match-

ing (LPM) or the Longest Preﬁx Classiﬁcation (LPC).

The different Interests with the same routable preﬁx

name issued by different end-users are forwarded un-

der the same FIB name-preﬁx. Furthermore, the Inter-

ests forwarded under the same FIB entry at one router

might be forwarded under different FIB entries at an-

other router, and therefore could be forwarded sepa-

rately.

To answer the second question, we examined the

NDN Pending Interest Table (NDN PIT) functional-

ities. As a core component, PIT is responsible for

keeping track of the awaiting Interest packets (Yi

et al., 2012; Yuan and Crowley, 2014). An Interest

is issued by one end-user but it could be aggregated

(becoming subsequent Interest) once or many times

at the downstream routers. Therefore, for a given

router within the network, a pending Interest could

be issued by one or many end-users. A returned data

object could satisfy one or multiple end-users. The

Interest aggregation feature combined with other fea-

tures (NDN in-network caching, adaptive forwarding

plane, ...) make the power of the NDN architecture. It

is important to notice that Interests to be aggregated

DCNET 2016 - International Conference on Data Communication Networking

should have the exact same preﬁx name and generally

this concerns the static content or a Content Delivery

Networks (CDN) content. Another aspect to consider

is the fact that under the same routable name (preﬁx-

name), we can ﬁnd a big number of different distinct

Interests issued by one or more than one end-user.

From the above discussion, it is clear that deﬁning

a ﬂow should be local to the router and for a speciﬁc

period of time (FIB update). In order to have a clear

idea, we model a Router R and its links with its neigh-

bor routers by a weighted directed graph G.

3.1 Router Model

A router R is connected to its K

ith

neighbor router by

a link L

with capacity C

, through the corresponding

face F

. The faces of the router R are completely con-

nected to each other with inﬁnite capacity links (Fig-

ure 1-a).

We model the router and its direct links as a di-

rected graph G = (F, L,C) (Figure 1-b) where:

• F is a set of vertices (Router faces and a virtual

central node R);

• L is the set of edges (Router links);

• C is a function whose domain is L, it is an assign-

ment of capacities to the edges (Router links in

terms of inputs and outputs bandwidth);

• The order of graph is k + 1 and the size of G is

|L(G)| = 2k;

• L

denotes the arc going from F

to R, represents

the downstream through the K

ith

face F

, with a

capacity C

;

• L

denote the arc going from R to F

, represents

the upstream through the K

ith

face F

, with a ca-

pacity C

;

• C

<= C

, is the capacity of bidirectional

link k connecting the router R and the K

ith

neigh-

bor router, through the face F

(Figure 1).

(a)

(b)

Figure 1: Router Model.

Deﬁnition 3.1. A Flow in NDN context is a set of In-

terests and their corresponding returned data objects

forwarded under the same FIB entry proper to a router

during a given period. These Interests may have been

requested from one or multiple faces and forwarded

through one or multiple faces (Figure 2).

Returned data under

The same FIB ent ry

Virtual End-User

Interests forwarded under

the same FIB entry

Figure 2: NDN Flow.

Example of NDN Flow: Let us consider an exam-

ple from Figure 2. In case the Interests were re-

ceived from face F

and F

, the FIB lookup mecha-

nism choses for them the same FIB entry with the cor-

responding FIB outgoing faces F

and F

. We con-

sider, with analogy to the connected oriented model:

• A virtual end-user source makes connections

(max. of six, three in our example) to the other

virtual end-user destination;

• The tuple (virtual end-user source, F

, R, F

, vir-

tual end-user destination) is a connection;

• If an Interest is forwarded through face F

and it

is received from face F

, the Interest will use the

resources l

through R and l

;

• The returned data will follow the reverse path l

through R and l

This set of Interests and their corresponding returned

data objects, are considered for this router as one

NDN ﬂow.

Parallel Multi-path Forwarding Strategy for Named Data Networking

Deﬁnition 3.2. In the presence of competing elastic

ﬂows, a Flow Assignment is fair in NDN context if

the assignment is weighted alpha-fair, where:

• Weights represent the normalized number of dis-

tinct awaiting sub-interests and the answered In-

terests from the Content Store (CS) as a function

of time.

• Distinct data object packets sizes and heteroge-

neous round-trip times (RT T s) are taken into con-

sideration.

3.2 Discussion

The fairness must be addressed across ﬂows, across

network and across time (Briscoe, 2007). The pro-

posed deﬁnition of NDN ﬂow and the fairness crite-

ria does not specify whether many ﬂows can serve a

common end-user. We claim in this paper a realis-

tic user-centric conception of fairness. The weighted

α-proportional fairness imposed between virtual end-

users ﬂows at each node ensures fairness between the

real end-users. The notion of friendliness imposed

by the emergence of the peer-to-peer networks in the

context of TCP/IP paradigm does not need to be ad-

dressed in the context of NDN. Taking the deﬁnitions

above, all the applications and sessions are friendly to

each other if the ﬂows are fairly tackled.

In (Yi, 2014), the authors present Fair Interest

Limiting (FIL). An NDN version of the fair queu-

ing mechanism used in the TCP/IP paradigm that suf-

fers from the same limitations in terms of fairness

(Briscoe, 2007). Besides, it does neither take the un-

predictable effect of content in-network caching and

the Interest aggregation feature nor the distinct data

object packets sizes and heterogeneous round-trip

times (RT T s) into consideration on the fairness evalu-

ation. At the opposite, weights presented in the above

deﬁnition come as an answer. The weights must be a

function of time to deal with the dynamic nature and

unpredictable changes in the network caused by the

effect of in-network caching and the Interest aggrega-

tion features.

4 PARALLEL MULTI-PATH

FORWARDING STRATEGY

Internet has a multi-path nature. This latter was not

exploited by the inherited Internet technologies al-

though its beneﬁts in terms of resilience and resource

usage efﬁciency (Qadir et al., 2015). The efforts of

deploying an Internet-scale multi-path protocol were

facing the end-to-end communication model. Ex-

changing path information and the large forwarding

tables lead to computational and storage overhead in

large-scale networks. In this context, scalability was

the black hole (Qadir et al., 2015; Baid et al., 2012).

In this section, we diagnose the problem, outline

the research problems and design considerations. Fur-

ther, we ﬁx the objectives and we propose a new for-

warding strategy.

4.1 Diagnosis: The Nature of the

Challenge

The future Internet is deﬁnitely multi-path (Qadir

et al., 2015). In this context, we propose the PMP-

FS which is a multi-path strategy that takes advantage

of NDN native support of parallel multi-path forward-

ing. It aims to serve end-users in a fair manner with-

out the need of identifying them while maximizing

network throughput. Here, it is signiﬁcant to men-

tion:

• The requested data objects in NDN may be cached

in different locations and in different periods of

time by the multi-homing mechanism and/or the

NDN in-caching mechanism. The whole data ob-

jects may be in one place or dispersed in different

places;

• The order of data packets arrival constitutes an-

other difference in forwarding packets between

NDN and TCP/IP paradigm. The order of data

packets arrival is very important in TCP/IP, which

is not the case in NDN. The returned data will be

recomposed for application use by end-users;

• Each ﬂow (session) aims at maximizing its

throughput;

• Resources are limited: the competition between

ﬂows as to be better served by forwarding them to

the set of the best possible face(s) is unavoidable;

• Preserving not only fairness between end-users

but also efﬁciency and scalability are a require-

ment;

• PMP-FS is to be executed at every NDN router;

• PMP-FS should respect the real time (line-speed)

constraint;

• PMP-FS should take the effect of content in-

network caching mechanism and the subsequent

awaiting Interests on the evaluation of fairness;

• Hop-by-hop Interest shaping mechanism to en-

sure the whole network stability is necessary;

• We assume that the ﬂows are elastic, the routing

plane is responsible for populating the FIB and we

DCNET 2016 - International Conference on Data Communication Networking

make no assumptions about whether the paths are

disjoint.

4.2 Guiding Policy: Dealing with the

Challenge

In (Yi et al., 2012; Yi et al., 2013; Yi, 2014), the

authors use multi-path feature that NDN offers for

recovering from failures (detect failure and retry al-

ternative path) and divert excess trafﬁc to other paths

which we call serial multi-path. This latter approach

ignores the competing nature of ﬂows. Oppositely,

parallel multi-path uses all or part of the available

paths in a controlled way to forward the Interests of a

ﬂow simultaneously over this set of available paths.

The PMP-FS splits trafﬁc and achieves fairness in

NDN context of the active ﬂows over multiple differ-

ent faces at each time slot. This increases reliability,

robustness and fault tolerance. The PMP-FS works as

follows:

• A hop-by-hop Interest shaping module proposed

in (Wang et al., 2013) is used as a congestion con-

trol mechanism. It is to be executed at every NDN

router to ensure the whole network stability. The

result of this module is the links input and out-

put capacities between the router at hand and its

neighbor routers;

• At every new FIB entry picked:

1. A new ﬂow queue created and set as inactive;

2. The ﬁrst group of Interests ( f Interests) is for-

warded immediately;

3. The ﬂow queue becomes active when the for-

warded Interests bring back data. The size of

data object is estimated;

4. Proportional Integral Controller Enhanced

(PIE) is used (Pan et al., 2013).

• At every data object packet received or cache hit:

1. Increment the per ﬂow counter.

• At every time slot:

1. the PMP-FS collects:

– The vector of per ﬂow counters;

– The information about the pending Interests

and the awaiting Interests for each active NDN

ﬂow and their corresponding FIB entries;

– The links input and output capacities: the re-

sult of the hop-by-hop Interest shaping capaci-

ties minus the reserved bandwidth for the end-

ing sessions (empty ﬂows queues).

2. Flow Assignment module is executed for the

active ﬂows queues: the result is a matrix A of

Flow Assignment;

3. Forwarding Decision module is executed:

– Taking as input:

∗ the matrix A of Flow Assignment of the last

step;

∗ the vector of the number of the pending In-

terests;

∗ the vector of estimated sizes of data objects

of the active ﬂows;

– The output is a matrix D of the number of In-

terests to forward for each ﬂow over each face;

• The matrix D is used to forward the incoming

Interests and for every Interest answered another

one from the same ﬂow is sent.

The time slot must be chosen in a manner that en-

able us to integrate the new ﬂows softly and elim-

inate the RT T variations effect on fairness. We

take 25 msec as the time slot which is about the

third of mean Internet RT T .

4.3 Action Plan: Carrying out the

Guiding Policy

NDN follows the pull-based model. Interests come

through one of the faces. Based on the FIB, these

Interests are forwarded through other face(s). The re-

turned data follows the reverse direction. The global

architecture of PMP-FS is depicted in Figure 3.

• At every new FIB entry picked:

– When a router receives an Interest and in case

the checking of the CS and the PIT results in

a negative response, the FIB is checked. At

this moment a new FIB entry is picked and the

a f ter receive Interest action is triggered.

– The variation in data object sizes is taken into

consideration by the per-ﬂow queuing (coun-

ters per name-space) and the ﬂow will be con-

sidered by the Flow Assignment module only

when it is active. In this case, we have the esti-

mated size of the data object packets. The latter

is smoothly updated at each received data ob-

ject (the be f ore satis f y Interest action is trig-

gered).

– Forwarding the ﬁrst group of Interests of a

newly created ﬂow over a set f of faces gives us

an opportunity to measure the performance and

rank the faces both dynamically and locally ( f

best faces ranked by the routing plane).

– Proportional Integral Controller Enhanced

(PIE) is used as an acceptance mechanism

by dropping incoming Interests and sending

pack NACKs based on a probability. Depar-

ture rate and the ﬂow queue length are used

Parallel Multi-path Forwarding Strategy for Named Data Networking

Figure 3: PMP-FS Global Architecture.

(Pan et al., 2013). When sending NACKs the

be f ore expire Interest action is triggered to up-

date the counters in measurements table.

• At every time slot (every 25 msec)

1. the PMP-FS collects:

– The information about the pending and the

awaiting Interests with their corresponding

FIB entries for each active NDN ﬂow and the

Interests that hit the cache. These will be used

as weights;

– The links input and output capacities: the re-

sult of the capacities of the hop-by-hop shap-

ing minus the reserved space for the ending

sessions (empty ﬂows queues);

2. The Flow Assignment module is executed for

the active ﬂows. The result is a matrix A of

Flow Assignment over the faces. The objective

of this module is to perform a controlled split-

ting of the active ﬂows over the faces while sat-

isfying the limited bandwidth constraints of the

outgoing and ingoing router links. In the next

section, the problem mathematical formulation

is presented.

3. The module of Forwarding Decision: The one-

to-one interdependence between Interests and

data objects called NDN ﬂow balance propri-

ety gives us the opportunity to carry out a con-

trolled splitting of the ﬂows over the faces by

only deciding where and how many Interests of

the active ﬂows to forward.

Taking as inputs the matrix of Flow Assignment

of the last step and the vector of estimated sizes

of the active ﬂows data objects, we get the ma-

trix D of the maximum number of Interests to

be forwarded for each ﬂow over each face.

The matrix D is used to forward the incoming

Interests, the next time slot and for every Inter-

est answered another one from the same ﬂow is

sent.

4.3.1 Network Model and Problem Formulation

To design The Flow Assignment module of the PMP-

FS, we use the model described previously (cf. sub-

section 3.1). We use the directed graph G = (F, L,C)

(cf. Figure 1).

For a better understanding, let us take the exam-

ple of the NDN ﬂow f w (see Figure 2): if Interests

of the same ﬂow enter through face F

and F

, their

corresponding FIB outgoing faces are face F

, F

and

. If an Interest of ﬂow f w is forwarded through face

and it is received from face F

. The latter Interest

would use the resources l

through R and l

. The

returned data object will follow the reverse path, l

through R and l

This ﬂow f w is competing with other ﬂows to

grab the max of bandwidth. The links capacities are

the result of the hop-by-hop Interest shaping mecha-

nism. The issue will be an optimization problem that

takes into consideration only the paths and seizes of

returned data objects.

The Flow Assignment module will decide the

splitting of active ﬂows (sessions) over the paths

while satisfying the limited bandwidth constraints. In

DCNET 2016 - International Conference on Data Communication Networking

other words, the module gives a solution of an opti-

mization problem for a weighted alpha fair ﬂow as-

signment. It takes into consideration the paths and

seizes of returned data objects only. The result of this

optimization will be used to decide on the number of

Interests of each ﬂow to be forwarded on each face.

Real time was always an issue when solving the

Flow Assignment and the Coordinated Multi-Path

Flow Control type of problems. The computation

scalability led to the use of off-line methods. The

authors in (McCormick et al., 2014) propose an al-

gorithm that could be executed in the range of mil-

liseconds and could be used as an on-line tool. It is

proposed as a Software Deﬁned Networking (SDN)

Trafﬁc Engineering tool based on the work presented

in (Voice, 2007). Oppositely to (McCormick et al.,

2014), we use the proposed algorithm in a distributed

manner to achieve weighted alpha-fairness ﬂow as-

signment at each router of the network. We apply the

algorithm in our proposed network model (cf. Fig-

ure 1).

The Flow Assignment optimization using

weighted alpha-fair utility functions is deﬁned as

follows:

Maximize

∑

s∈S

1−α

1 − α

(2)

Sub ject to







∑

p∈l

≤ C

∑

p∈s

= x

(3)

f or w > 0, α 6= 1, x > 0, y > 0, l ∈ L

The resulting updating rules, following (Mc-

Cormick et al., 2014; Kelly et al., 2008; Voice, 2007),

using the same notations as in Table 1 are:

= ((

s(p)

)

∑

l∈p

)

1−q

s(p)

(4)

(t + 1) = µ

(t) +

1−q

(t)

∑

p∈l

(t)−c

(5)

(t + 1) = x

(t) +

1−q

2(α+q−1)

(t)

∑

p∈s

(t)

−x

(t)

(6)

The main advantage of this formulation is that the

updating rules can be implemented in parallel. The

(t) is the shadow price of the link l at iteration t of

the path p.

In the sequel, we present some preliminary results.

Furthermore, we discuss implementation and evalua-

tion considerations.

5 PERFORMANCE EVALUATION

The general structure of the module describing how

we can implement the formulation described in sub-

subsection 4.3.1 is shown in algorithm 1.

Algorithm 1: Flow Assignment Module.

Input: G = (F,L,C), Weights W [ f l].

1 begin for All Flows s do

2 for All Sub Flows of s do

3 for All Sub Flows Y

4 y

← ((

s(p)

)

∑

l∈p

)

1−q

s(p)

5 repeat

6 for All links l do

7 µ

← µ

1−q

∑

p∈l

−c

8 for All Aggregate Flows X

9 x

← x

1−q

2(α+q−1)

∑

p∈s

−x

10 Y

old ← y

11 for All Sub Flows Y

12 y

← ((

s(p)

)

∑

l∈p

)

1−q

s(p)

13 until (|y

−Y

old| < 10

−5

)

14 return Matrix of Flow Assignment A

5.1 Preliminary Results

An implementation of algorithm 1 in C++ on an Intel

Core i5-3550 (6M Cache, 3.3 GHz) processor gives

the results shown in Figure 4.

0 5 0 0 1 0 0 0 1 5 0 0 2 0 0 0 2 5 0 0

2 0 0 0

4 0 0 0

6 0 0 0

8 0 0 0

E x e c u t i o n T i m e ( s )

N u m b e r o f F l o w s

N u m b e r o f F a c e s = 1 0 0

α= 4

q = 0 . 7 5

Figure 4: Execution times per number of ﬂows without par-

allelism.

The execution time grows exponentially with the

number of ﬂows, which explains the real-time issue

that was previously mentioned. In Figure 5, we stud-

ied the effect of number of faces on the execution

time. Obtained results show that the number of faces

does not affect the execution time.

The Field-programmable Gate Array (FPGA) im-

plementation of algorithm 1 on BT 21CN network

Parallel Multi-path Forwarding Strategy for Named Data Networking

0 5 0 1 0 0 1 5 0 2 0 0 2 5 0

5 0

1 0 0

1 5 0

2 0 0

2 5 0

3 0 0

N u m b e r o f F l o w s = 5 0 0

α= 4

q = 0 . 7 5

T i m e o f E x e c u t i o n ( s )

N u m b e r o f F a c e s

Figure 5: Execution times per number of faces without par-

allelism.

comprising 106 nodes and 234 links was done by

(McCormick et al., 2014) on a Virtex 7 FPGA. The

simulation was run with more than ten thousand

(10000) ﬂows in less than two (2) msec (McCormick

et al., 2014). If we have a router with hundred (100)

faces, following our model we would have one hun-

dred and one (101) nodes and two hundred (200) links

network. The RT T in todays Internet is around 80 ms

on average (Perino and Varvello, 2011). We take

25 msec as the time slot, which is about the third of In-

ternet RT T average. With the latter frequency of exe-

cution of the Flow Assignment module (three times in

one RT T round), PMP-FS can integrate the new ﬂows

softly and eliminate the RT T variations effect on fair-

ness. This proves the computation time scalability of

our PMP-FS.

5.2 Discussion

The actual conﬁguration of the FPGA implementa-

tion of (McCormick et al., 2014) supports 512 links,

32000 ﬂows and 96K paths. Using our model, the

latter conﬁguration is equivalent to a router with 256

faces and 32000 NDN ﬂows.

The number of active ﬂows in a core router is in

the order of 10

(Mills et al., 2010), which makes the

PMP-FS an excellent choice for use on Internet scale

network. But it is signiﬁcant to mention that one has

to design for the future. We argue, besides the techni-

cal advances in the future that would make the mod-

ule limits higher, our proposal is conﬁgurable. It is

in the same spirit as (Afanasyev and Wang, 2015).

The use of hierarchical names by NDN made our pro-

posal ﬂexible to the level of granularity. Maintaining

per-ﬂow states information is undesirable in large net-

works but in the case of PMP-FS one has to see it as

a class-level or ﬂow aggregation with different gran-

ularities. Setting the degree of granularity (the level

of the FIB entry in the name tree to consider by the

Flow Assignment Module) depend on the role of the

router in the network (Edge or Core Router). The lat-

ter ﬂexibility makes the PMP-FS adaptable to support

any exponential growth of Internet. In this context, it

is worth highlighting that:

• The choice of the degree of granularity combined

with the weights of NDN ﬂows makes PMP-FS

able to handle any number of ﬂows. The fairness

would not be affected and can be controlled by the

choice of the degree of granularity;

• The degree of granularity is used to reduce the

per-ﬂow state managed by each router and to con-

trol the processing time and storing space;

• The weights of NDN ﬂows can be used to support

differentiated bandwidth allocation (based on pri-

ority);

• The in and out link capacities used by the Assign-

ment Flow Module can be used and conﬁgured to

leave room in the links for QoS guaranteed appli-

cations.

• The packets (Interests and Data Objects) are for-

warded in line speed.

6 CONCLUSION AND FUTURE

WORK

In this paper, we have proposed a new Parallel Multi-

Path Forwarding Strategy (PMP-FS) that takes into

consideration NDN in-network caching and NDN In-

terest aggregation features to achieve weighted alpha

fairness among different NDN ﬂows. In one hand, the

proposed PMP-FS is a general framework for han-

dling fairness and throughput (by setting the tuning

parameters (e.g. α and q)). In the other hand, it

can handle all type of application needs (delay, band-

width, ...) and communication protocols (over TCP,

Ethernet, ...). Besides it can be used to implement

differentiated services by routable name-preﬁx pro-

viding quality of service (QoS). We do believe that

the major drawback of the proposed solution is the

need of an extra hardware. However, we do believe

also that it will add value of robustness and efﬁciency

in the future Internet architecture.

The evaluation of the PMP-FS in the ndnSIM 2.1,

an open-source simulator for NDN (Afanasyev et al.,

2012; Mastorakis et al., 2015), is a work-in-progress.

In this context, it is worth pointing out that the for-

warding strategies implemented in the ndnSIM are

special cases of the proposed PMP-FS strategy:

• Best Route Strategy: PMP-FS with one best link;

DCNET 2016 - International Conference on Data Communication Networking

• Multi-cast Strategy: PMP-FS with all upstreams

indicated by FIB entry;

• Client Control Strategy: PMP-FS testing speciﬁc

Interests and allowing them to be forwarded to

preconﬁgured outgoing faces.

REFERENCES

Afanasyev, A., Burke, J., Zhang, L., Claffy, K., Wang,

L., Jacobson, V., Crowley, P., Papadopoulos, C.,

and Zhang, B. (2014). Named Data Networking.

ACM SIGCOMM Computer Communication Review,

44(3):66–73.

Afanasyev, A., Moiseenko, I., and Zhang, L. (2012).

ndnSIM: NDN simulator for NS-3. Technical report,

NDN-0005.

Afanasyev, A., Shi, J., Zhang, B., Zhang, L., Moiseenko,

I., Yu, Y., Shang, W., Huang, Y., Abraham, J. P.,

Dibenedetto, S., Fan, C., Pesavento, D., Grassi, G.,

Pau, G., Zhang, H., Song, T., Abraham, H. B., Crow-

ley, P., Amin, S. O., Lehman, V., and Wang, L.

(2015a). NFD Developer ’ s Guide. Technical report,

NDN-0021.

Afanasyev, A. and Wang, L. (2015). Map-and-Encap for

Scaling NDN Routing. Technical report, NDN-0004.

Afanasyev, A., Yi, C., Wang, L., Zhang, B., and Zhang,

L. (2015b). SNAMP: Secure Namespace Mapping to

Scale NDN Forwarding. IEEE Global Internet Sym-

posium (GI 2015).

Baid, A., Vu, T., and Raychaudhuri, D. (2012). Comparing

alternative approaches for networking of named ob-

jects in the future Internet. In Computer Communica-

tions Workshops (INFOCOM WKSHPS), 2012 IEEE

Conference on Emerging Design Choices in Name-

Oriented Networking, pages 298–303.

Briscoe, B. (2007). Flow rate fairness: Dismantling a re-

ligion. ACM SIGCOMM Computer Communication

Review, 37(2):63–74.

Hahne, E. L. (1991). Round-robin scheduling for max-min

fairness in data networks. Selected Areas in Commu-

nications, IEEE Journal on, 9(7):1024–1039.

Jacobson, V., Smetters, D. K., Briggs, N. H., Plass, M. F.,

Stewart, P., Thornton, J., and Braynard, R. L. (2009).

VoCCN: Voice-over Content-Centric Networks. Pro-

ceedings of the 2009 workshop on Re-architecting the

internet, pages 1–6.

Kelly, F. (1997). Charging and rate control for elastic traf-

ﬁc. European Transactions on Telecommunications,

8(1):33–37.

Kelly, F., Raina, G., and Voice, T. (2008). Stability and fair-

ness of explicit congestion control with small buffers.

ACM SIGCOMM Computer Communication Review,

38(3):51.

Low, S. H. (2003). A duality model of TCP and queue man-

agement algorithms. IEEE/ACM Transactions on net-

working, 11(4):525–536.

Mastorakis, S., Afanasyev, A., Moiseenko, I., and Zhang,

L. (2015). ndnSIM 2 . 0 : A new version of the NDN

simulator for NS-3. Technical report, NDN-0028.

Mazumdar, R., Mason, L. G., and Douligeris, C. (1991).

Fairness in network optimal ﬂow control: optimality

of product forms. Communications, IEEE Transac-

tions on, 39(5):775–782.

McCormick, B., Kelly, F., Plante, P., Gunning, P., and

Ashwood-Smith, P. (2014). Real time alpha-fairness

based trafﬁc engineering. Proceedings of the third

workshop on Hot topics in software deﬁned network-

ing - HotSDN ’14, 1(1):199–200.

Mills, K., Filliben, J., Cho, D., Schwartz, E., and Genin, D.

(2010). Study of Proposed Internet Congestion Con-

trol Algorithms. Technical report, NIST Special Pub-

lication 500-282.

Mo, J. and Walrand, J. (2000). Fair end-to-end Window-

based Congestion Control. IEEE/ACM Trans. Netw.,

8(5):556–567.

Narayanan, A. and Oran, D. (2015). Ndn and Ip Routing

Can It Scale ? In Proposed Information-Centric Net-

working Research Group (ICNRG), Side meeting at

IETF-82, Taipei.

Pan, R., Natarajan, P., Piglione, C., Prabhu, M. S., Subra-

manian, V., Baker, F., and VerSteeg, B. (2013). PIE: A

lightweight control scheme to address the bufferbloat

problem. In 14th International Conference on High

Performance Switching and Routing (HPSR), IEEE,

pages 148–155, Taipei, Taiwan.

Perino, D. and Varvello, M. (2011). A reality check for

content centric networking. In Proceedings of the

ACM SIGCOMM workshop on Information-centric

networking - ICN ’11, page 44, New York, USA.

ACM Press.

Qadir, J., Ali, A., Yau, K.-l. A., Sathiaseelan, A., and

Crowcroft, J. (2015). Exploiting the power of mul-

tiplicity: a holistic survey of network-layer multipath.

arXiv:1502.02111v1 [cs.NI], pages 1–35.

Song, T., Yuan, H., Crowley, P., and Zhang, B. (2015). Scal-

able Name-Based Packet Forwarding: From Millions

to Billions. In Proceedings of the Second Interna-

tional Conference on Information-Centric Network-

ing, ICN ’15, pages 19–28, New York, USA. ACM.

Voice, T. (2007). Stability of multi-path dual congestion

control algorithms. IEEE/ACM Transactions on Net-

working, 15(6):1231–1239.

Wang, Y. (2013). Caching, Routing and Congestion Control

in a Future Information-Centric Internet. PhD thesis,

North Carolina State University.

Wang, Y., Rozhnova, N., Narayanan, A., Oran, D., and

Rhee, I. (2013). An improved hop-by-hop interest

shaper for congestion control in named data network-

ing. ACM SIGCOMM Computer Communication Re-

view, 43(4):55–60.

Yi, C. (2014). Adaptive forwarding in named data network-

ing. PhD thesis, The University Of Arizona.

Yi, C., Afanasyev, A., Moiseenko, I., Wang, L., Zhang, B.,

and Zhang, L. (2013). A Case for Stateful Forwarding

Plane. Computer Communications, 36(7):779–791.

Parallel Multi-path Forwarding Strategy for Named Data Networking

Yi, C., Afanasyev, A., Wang, L., Zhang, B., and Zhang, L.

(2012). Adaptive forwarding in named data network-

ing. ACM SIGCOMM Computer Communication Re-

view, 42(3):62.

Yuan, H. and Crowley, P. (2014). Scalable Pending Interest

Table design: From principles to practice. Proceed-

ings - IEEE INFOCOM, pages 2049–2057.

Yuan, H., Song, T., and Crowley, P. (2012). Scalable NDN

Forwarding: Concepts, Issues and Principles. 21st

International Conference on Computer Communica-

tions and Networks (ICCCN), pages 1–9.

Zhang, L., Estrin, D., Burke, J., Jacobson, V., Thorton,

J. D., Smetters, D. K., Zhang, B., Tsudik, G., Claffy,

K., Krioukov, D., Massey, D., Papadopoulos, C., Ab-

delzaher, T., Wang, L., Crowley, P., and Yeh, E.

(2010). Named Data Networking. Technical report,

NDN-0001.

DCNET 2016 - International Conference on Data Communication Networking