Previous | 
Home
Section: New Results
Towards Data-Centric Networking
Participants : Chadi Barakat, Damien Saucez, Jonathan Detchart, Mohamed Ali Kaafar, Ferdaouss Mattoussi, Marc Mendonca, Xuan-Nam Nguyen, Vincent Roca, Thierry Turletti.
-
Delay Tolerant Networks (DTNs) stand for wireless networks where disconnections may occur frequently. In order to achieve data delivery in such challenging environments, researchers have proposed the use of store-carry-and-forward protocols: there, a node may store a message in its buffer and carry it along for long periods of time, until an appropriate forwarding opportunity arises. Multiple message replicas are often propagated to increase delivery probability. This combination of long-term storage and replication imposes a high storage and bandwidth overhead. Thus, efficient scheduling and drop policies are necessary to: (i) decide on the order by which messages should be replicated when contact durations are limited, and (ii) which messages should be discarded when nodes' buffers operate close to their capacity.
We worked on an optimal scheduling and drop policy that can optimize different performance metrics, such as the average delivery rate and the average delivery delay. First, we derived an optimal policy using global knowledge about the network, then we introduced a distributed algorithm that collects statistics about network history and uses appropriate estimators for the global knowledge required by the optimal policy, in practice. At the end, we are able to associate to each message inside the network a utility value that can be calculated locally, and that allows to compare it to other messages upon scheduling and buffer congestion. Our solution called HBSD (History Based Scheduling and Drop) integrates methods to reduce the overhead of the history-collection plane and to adapt to network conditions. The first version of HBSD and the theory behind have been published in 2008. A recent paper [27] provides an extension to a heterogenous mobility scenario in addition to refinements to the history collection algorithm. An implementation is proposed for the DTN2 architecture as an external router and experiments have been carried out by both real trace driven simulations and experiments over the SCORPION testbed at the University of California Santa Cruz. We refer to the web page of HBSD for more details http://planete.inria.fr/HBSD_DTN2/ .
HBSD in its current version is for point-to-point communications. Another interesting schema is to consider one-to-many communications, where requesters for content express their interests to the network, which looks for the content on their behalf and delivers it back to them. Along the main ideas of HBSD, we worked on a content optimal-delivery algorithm, CODA, that distributes content to multiple receivers over a DTN. CODA assigns a utility to each content item published in the network; this value gauges the contribution of a single content replica to the network’s overall delivery-rate. CODA performs buffer management by first calculating the delivery-rate utility of each cached content-replica and then discarding the least-useful item. When an application requests content, the node supporting the application will look for the content in its cache. It will immediately deliver it to the application if the content is stored in memory. In case the request cannot be satisfied immediately, the node will store the pending request in a table. When the node meets another device, it will send the list of all pending requests to its peer; the peer device will try to satisfy this list by sending the requester all the matching content stored in its own buffer. A meeting between a pair of devices might not last long enough for all requested content to be sent. We address this problem by sequencing transmissions of data in order of decreasing delivery-rate utility. A content item with few replicas in the network has a high delivery rate utility; these items must be transmitted first to avoid degrading the content delivery-rate metric. The node delivers the requested content to the application as soon as it receives it in its buffer. We implement CODA over the CCNx protocol, which provides the basic tools for requesting, storing, and forwarding content. Detailed information on CODA and the implementation work carried out herein can be found in [76] .
Naming and Routing in Content Centric Networks
Content distribution prevails in todays Internet and content oriented networking proposes to access data directly by their content name instead of their location, changing so the way routing must be conceived. We proposed a routing mechanism that faces the new challenge of interconnecting content-oriented networks. Our solution relies on a naming resolution infrastructure that provides the binding between the content name and the content networks that can provide it. Content-oriented messages are sent encaspulated in IP packets between the content-oriented networks. In order to allow scalability and policy management, as well as traffic popularity independence, binding requests are always transmitted to the content owner. The content owner can then dynamically learn the caches in the network and adapt its binding to leverage the cache use.
The work done so far is related to routing between content-oriented networks. We are starting an activity on how to provide routing inside a content network. To that aim, we are investigating on the one hand probabilistic routing and, on the other hand, deterministic routing and possible extension to Bellman-Ford techniques. In addition to routing, we are investigating the problem of congestion in content-oriented networks. Indeed, in this new paradigm, congestion must be controlled on a per-hop basis, as opposed to the end-to-end congestion control that prevails today. We think that we can combine routing and congestion control to optimize resource consumption. Finally, we are studying the implications of using CCN from an economical perspective. See [100] for more details.
-
Content-centric networking (CCN) is a new paradigm to better handle contents in the future Internet. Under the assumption that CCN networks will deploy a similar congestion control mechanism than in today's TCP/IP (i.e., AIMD), we built an analytical model of the bandwidth sharing in CCN based on the “square-root formula of TCP”. With this model we can compare CCN download performance to what users get today. We consider different factors such as the way CCN routers are deployed, the popularity of contents, or the capacity of links and observe that when AIMD is used in a CCN network less popular content throughput is massively penalised whilst the individual gain for popular content is negligible. Finally, the main advantage of using CCN is the decrease of load at the server side. Our observations advocate the necessity to clearly define the notion of fairness in CCN and to design a proper congestion control to avoid less popular contents to become hardly accessible in tomorrow's Internet.
Our results [75] clearly point to a fairness issue if AIMD is used with CCN. Indeed, combining blindly AIMD and CCN can severely worsen the download throughput of less popular contents with respect to the today’s Internet due to subtle interactions with in-network caching strategies. The way cache memories are distributed within chain topologies has been investigated too, showing that for small and heterogeneous cache spaces, placing the biggest caches close to clients improves performance due to a smaller RTT on average. On the other hand, CCN can significantly reduce the load at the server side independently of the cache allocation strategy. Our findings advocate the urge of clearly defining the notion of fairness in CCN and designing congestion control algorithms able to limit the unfairness observed between contents of different popularities. The work is currently used within the IRTF ICNRG research group in order to motivate and define an appropriate congestion control mechanism for information centric networks like CCN. Moreover, we are currently validating the analytical results with an implementation of CCN where we can evaluate how much our model deviates from the reality when contents are of various size or small. The implementation will also be a support to test different congestion control mechanism.
CCN to enable profitable collaborative OTT services
The ubiquity of broadband Internet and the proliferation of connected devices like laptops, tablets, or TV result in a high demand of multimedia content such as high definition video on demand (VOD) for which the Internet has been poorly designed with the Internet Protocol (IP). Information-Centric Networking and more precisely Content Centric Networking (CCN) overtake the limitation of IP by considering content as the essential element of the network instead of the topology. CCN and its content caching capabilities is particularly adapted to Over-The-Top (OTT) services like Netflix, Hulu, Xbox Live, or YouTube that distribute high-definition multimedia content to millions of consumers, independently of their location. However, bringing content as the most important component of the network implies fundamental changes in the Internet and the transition to a fully CCN Internet might take a long time. Despite this transition period where CCN and IP will co-exist, we have shown that OTT service providers and consumers have strong incentives for migrating to CCN. We also propose a transition mechanism based on the Locator/Identifier Separation Protocol (LISP) [28] that allows the provider to track the demands from its consumers even though they do not download the contents from another consumers instead of the producer itself.
CCN, compared to IP, provides better security and performance. This last point is very interesting for OTT service providers that deliver multimedia content where performance is a key factor for the adoption of the service by consumers. With CCN, the content can be retrieved from the caches in the different CCN islands, instead of always being delivered by the content publisher. As a result, content retrieval is faster for the consumer and the operational cost of the publisher is reduced. Moreover, as the content is cached by the consumers and because the consumer can provide the content to other consumers, the overall performance increases with the number of consumers instead of decreasing as it is the case in IP today where the content is delivered by the hosting server. This property is particularly interesting because it dampens the effect of flash crowds which are normally very costly for OTT service providers as they have to provision their servers and networks to support them. Using CCN with caching at the consumers has then a direct impact on the profit earned by the OTT service provider as its costs are reduced. However, to benefit from the caching capabilities of consumers, the producer must propose real incentives to its consumers to collaborate and cache the content. To understand how incentives can be provided, it is necessary to remember that content in OTT is provided either freely to the consumer or in exchange of a fee. When the content is provided freely, the incomes for the publisher are ensured by advertisements dispersed in the content (e.g., banner, commercial interruptions...). A consumer has incentives to collaborate with the system if it receives some sort of discount, expressed in advertisement reduction or fee reduction. On the one hand, the discount has a cost for the publisher as its revenues will be reduced. On the other hand, the collaboration from its consumers reduces its operational costs. Hence, the publisher must determine the optimal discount, such that it maximises its profit. The situation for the consumer is the exact opposite: its costs are increasing because it is providing content to other consumers but its revenues also increase as it receives a discount on its expenses. We have determined the conditions to respect when deploying OTT with loosely collaborative consumers [99] . We currently refine the results using game theory.
Software-Defined Networking in Heterogeneous Networked Environments
Software-Defined Networking (SDN) has been proposed as a way to facilitate network evolution by allowing networks and their infrastructure to be programmable. In the context of the COMMUNITY associated team with University of California Santa Cruz (see URL http://inrg.cse.ucsc.edu/community/ ) , we are studying the potential of SDN to facilitate the deployment and management of new architectures and services in heterogeneous environments. In particular, we focus on the fundamental issues related to enabling SDN in infrastructure-less/decentralized networked environments and we use OpenFlow as our target SDN platform. Our plan is to develop a hybrid SDN framework that strikes a balance between a completely decentralized approach like Active Networking and a centralized one such as OpenFlow˜[58] .
We are also currently evaluating the efficiency of SDN for optimizing caching in content-centric networks. CCN advocates in-network caching, i.e., to cache contents on the path from content providers to requesters. Akthough this on-path caching achieves good overall performance, we have shown that this strategy is far from being the optimal inside a domain. On this purpose, we proposed the notion of off-path caching by allowing deflection of the most popular traffic off the optimal path towards off-path caches available across the domain[100] . Off-path caching improves the global hit ratio and permits to reduce the peering links' bandwidth usage. We are now investigating whether SDN functionalities can be used to implement this optimal caching technique, in particular to identify of the most popular contents, and to configure deflection mechanisms within routers˜[94] .
Application-Level Forward Error Correction Codes (AL-FEC) and their Applications to Broadcast/Multicast Systems
With the advent of broadcast/multicast systems (e.g., 3GPP MBMS services), large scale content broadcasting is becoming a key technology. This type of data distribution scheme largely relies on the use of Application Level Forward Error Correction codes (AL-FEC), not only to recover from erasures but also to improve the content broadcasting scheme itself (e.g., with FLUTE/ALC).
Our LDPC-Staircase codes, that offer a good balance in terms of performance, have been included as the primary AL-FEC solution for ISDB-Tmm (Integrated Services Digital Broadcasting, Terrestrial Mobile Multimedia), a Japanese standard for digital television (DTV) and digital radio, with a commercial service that started in April 2012. This is the first adoption of these codes in an international standard. These codes, along with our FLUTE/ALC software, are now part of the server and terminal protocol stack: http://www.rapidtvnews.com/index.php/2012041721327/ntt-data-mse-and-expways-joint-solution-powers-japanese-mobile-tv-service.html .
This success has been made possible, on the one hand, by major efforts in terms of standardization within IETF: the RFC 5170 (2008) defines the codes and their use in FLUTE/ALC, a protocol stack for massively scalable and reliable content delivery services, an active Internet-Draft published last year describes the use of these AL-FEC codes in FECFRAME, a framework for robust real-time streaming applications, and recent Internet-Drafts [91] [92] define the GOE (Generalized Object Encoding) extension of LDPC-Staircase codes for UEP (Unequal Erasure Protection) and file bundle protection services.
This success has also been made possible, on the other hand, by our efforts in terms of design and evaluation of two efficient software codecs for LDPC-Staircase codes. One of them is distributed in open-source, as part of our OpenFEC project (http://openfec.org), a unique initiative that aims at promoting open and free AL-FEC solutions. The second one, a highly optimized version with improved decoding speed and reduced memory requirements, is commercialized through an industrial partner, Expway.
Since May 2012, along with the Expway French company, we are proposing the Reed-Solomon + LDPC-Staircase codes for the 3GPP-eMBMS call for technology, as a candidate for next generation AL-FEC codes for multimedia services. We have shown that these codes offer very good erasure recovery capabilities, in line with 3GPP requirements, and extremely high decoding speeds, usually significantly faster than that of the other proposals. The final decision is expected for end of January 2013. In any case we have once again showed that these codes provide very good performance, often ahead of the competitors, and an excellent balance between several technical and non technical criteria.
Finally our activities in the context of the PhD of F. Mattoussi include the design, analysis and improvement of GLDPC-Staircase codes, a "Generalized" extension to LDPC-Staircase codes. We have shown in particular that these codes: (1) offer small rate capabilities, i.e. can produce a large number of repair symbols 'on-the-fly', when needed; (2) feature high erasure recovery capabilities, close to that of ideal codes. Therefore they offer a nice opportunity to extend the field of application of existing LDPC-Staircase codes (IETF RFC 5170), while keeping backward compatibility (i.e. LDPC-Staircase "codewords" can be decoded with a GPLDPC-Staircase codec). More information is available in [56] [57] [55] .
Unequal Erasure Protection (UEP) and File bundle protection through the GOE (Generalized Object Encoding) scheme
This activity has been initiated with the PostDoc work of Rodrigue IMAD. It focuses on Unequal Erasure Protection capabilities (UEP) (when a subset of an object has more importance than the remaing) and file bundle protection capabilities (e.g. when one want to globally protect a large set of small objects).
After an in-depth understanding of the well-known PET (Priority Encoding Technique) scheme, and the UOD for RaptorQ (Universal Object Delivery) initiative of Qualcomm, which is a realization of the PET approach, we have designed the GOE FEC Scheme (Generalized Object Encoding) alternative. The idea, simple, is to decouple the FEC protection from the natural object boundaries, and to apply an independant FEC encoding to each "generalized object". The main difficulty is to find an appropriate signaling solution to synchronize the sender and receiver on the exact way FEC encoding is applied. In [91] we show this is feasible, while keeping a backward compatibility with receivers that do not support GOE FEC schemes. Two well known AL-FEC schemes have also been extended to support this new approach, with very minimal modifications, namely Reed-Solomon and LDPC-Staircase codes [92] , [91] .
During this work, we compared the GOE and UOD/PET schemes, both from an analytical point of view (we use an N-truncated negative binomial distribution to that purpose) and from an experimental, simulation based, point of view [64] . We have shown that the GOE approach, by the flexibility it offers, its simplicity, its backward compatibility and its good recovery capabilities (under finite of infinite length conditions), outperforms UOD/PET for practical realizations of UEP/file bundle protection systems. See also http://www.ietf.org/proceedings/81/slides/rmt-2.pdf .
Application-Level Forward Error Correction Codes (AL-FEC) and their Applications to Robust Streaming Systems
AL-FEC codes are known to be useful to protect time-constrained flows. The goal of the IETF FECFRAME working group is to design a generic framework to enable various kinds of AL-FEC schemes to be integrated within RTP/UDP (or similar) data flows. Our contributions in the IETF context are three fold. First of all, we have contributed to the design and standardization of the FECFRAME framework, now published as a Standards Track RFC6363.
Secondly, we have proposed the use of Reed-Solomon codes (with and without RTP encapsulation of repair packets) and LDPC-Staircase codes within the FECFRAME framework: [85] for Reed-Solomon and [88] for LDPC-Staircase. Both documents are close to being published as RFCs.
Finally, in parallel, we have started an implementation of the FECFRAME framework in order to gain an in-depth understanding of the system. Previous results showed the benefits of LDPC-Staircase codes when dealing with high bit-rate real-time flows.
A second type of activity, in the context of robust streaming systems, consisted in the analysis of the Tetrys approach. Tetrys is a promising technique that features high reliability while being independent from RTT, and performs better than traditional block FEC techniques in a wide range of operational conditions.
A new File Delivery Application for Broadcast/Multicast Systems
FLUTE [95] has long been the one and only official file delivery application on top of the ALC reliable multicast transport protocol. However FLUTE has several limitations (essentially because the object meta-data are transmitted independently of the objects themselves, in spite of their inter-dependency), features an intrinsic complexity, and is only available for ALC.
Therefore, we started the design of FCAST, a simple, lightweight file transfer application, that works both on top of both ALC and NORM [82] . This work is carried out as part of the IETF RMT Working Group, in collaboration with B. Adamson (NRL). This document has passed WG Last Call and is currently considered by IESG.
Security of the Broadcast/Multicast Systems
Sooner or later, broadcasting systems will require security services. This is all the more true as heterogeneous broadcasting technologies are used, some of them being by nature open, such as WiFi networks. Therefore, one of the key security services is the authentication of the packet origin and the packet integrity check. To that purpose, we have specified the use of simple authentication and integrity schemes (i.e., group MAC and digital signatures) in the context of the ALC and NORM protocols and the standard is now published as IETF RFC 6584 [98] .
High Performance Security Gateways for High Assurance Environments
This work focuses on very high performance security gateways, compatible with 10Gbps or higher IPsec tunneling throughput, while offering a high assurance thanks in particular to a clear red/black flow separation. In this context we have studied last year the feasibility of high-bandwidth, secure communications on generic machines equipped with the latest CPUs and General-Purpose Graphical Processing Units (GPGPU).
The work carried out in 2011-2012 consisted in setting up and evaluating the high performance platform. This platform heavily relies on the Click modular TCP/IP protocol stack implementation, which turned out to be a key enabler both in terms of specialization of the stack and parallel processing. Our activities also consisted in analyzing the PMTU discovery aspect since it is a critical factor in achieving high bandwidths. To that goal we have designed a new approach for qualifying ICMP blackholes in the Internet, since PMTUD heavily relies on ICMP [51] .