ICP: Design and Evaluation of an Interest Control Protocol for Content-Centric Networking

Transcription

1 ICP: Design and Evaluation of an Interest Control Protocol for Content-Centric Networking Giovanna Carofiglio, Massimo Gallo, Luca Muscariello Bell Labs, Alcatel-Lucent, Orange Labs, France Telecom, Abstract Content-centric networking (CCN) brings a paradigm shift in the present Internet communication model by addressing named-data instead of host locations. With respect to TCP/IP, the transport model is connectionless with a unique endpoint at the receiver, driving a retrieval process natively point to multi-point. Another salient feature of CCN is the possibility to embed storage capabilities into the network, adding a new dimension to the transport problem. The focus of this work is on the design of a receiver-driven Interest control protocol for CCN, whose definition, to the best of our knowledge, still lacks in literature. ICP realizes a window-based Interest flow control, achieving full efficiency and fairness under proper parameters setting. In this paper, we provide an analytical characterization of average rate, expected data transfer delay and queue dynamics in steady state on a single and multi-bottleneck network topology. Our model accounts for the impact of on-path caches. Protocol performance is also assessed via packet-level simulations and design guidelines are drawn from previous analysis. I. INTRODUCTION The move towards a future Internet is today hindered by the mismatch between the host-oriented model, at the foundation of the current network architecture, and the dominant content-oriented usage, centered on data dissemination and retrieval. With the spread of data-centric services, the need for a content-aware infrastructure has been addressed by application-layer solutions like CDNs, PP overlays and HTTP proxies, deployed on top of the current Internet infrastructure. However, such solutions suffer from a number of limitations in terms of efficient resources utilization, scalability and management cost, mainly because they are engineered as overlay service-dedicated infrastructures at application layer. Content-centric networking (CCN) proposals have then emerged to redesign the Internet architecture around named data, in order to natively realize content dissemination at networking level. Pioneered by Van Jacobson [] but also previously advocated in [], the idea of a network architecture centered on named data has recently gained momentum in the research community and originated various research efforts (see [], for a list of current projects). As a shared principle of CCN proposals, content is uniquely identified, addressed and retrieved by its name independently from its location, and storage capabilities are embedded into the network. Previous work on content-centric networks have mainly focused on the global architecture design [], [], [], [5], while less effort has been devoted to the study of the interaction between caching and transport mechanisms in such architectures and notably, on the definition of a CCN flow control protocol. CCN caching benefits have been explored independently. Somaya et al. [] recently analyze packet-level caching feasibility in routers at line-speed, while Lee et al. [7] consider the benefits of CCN in-network storage in terms of energy efficiency with respect to traditional distribution architectures. Also, Carofiglio et al. [8] show the role played by storage management in CCN by means of experimental evaluation. A first analytical characterization of transport performance and of its interaction with packet-level caching dynamics is provided in [9], with no bandwidth limitation. In [], the modeling framework is extended to the case of limited bandwidth (and storage) resources assuming perfect fair bandwidth sharing among flows. Our work aims at complementing existing CCN designs by focusing on the definition and the analysis of a suitable and reliable transport protocol. The contribution of this paper is threefold: (i) We design a window-based Interest flow control protocol driven by an AIMD (Additive Increase Multiplicative Decrease) mechanism to regulate the Interest rate at the receiver. (ii) We analyze the designed protocol on a single and multibottleneck link scenario, by accounting for the interaction with in-network caching, building upon our previous storage sharing model in [9]. Our analysis also proves that ICP guarantees optimal fair and efficient bandwidth sharing, as described in []. (iii) We carry out packet-level simulations to corroborate analytical findings and give guidelines on parameter setting. The rest of the paper is organized as follows: in Sec.II we describe the system and summarize model assumptions. Sec.III describes ICP design, while Secc.IV-VI report the analysis respectively in a single and multi-bottleneck scenario. Guidelines on ICP parameters setting are provided in Sec.VII-A, whereas numerical results are gathered in Sec.VII-B. Finally, Sec.VIII concludes the paper. II. SYSTEM DESCRIPTION AND MODEL ASSUMPTIONS In this paper we primarily focus on the content-centric networking (CCN/NDN) proposal by Parc [] (see also []), though the modeling framework has broader applicability in the context of Information Centric Networking (ICN []) and more generally receiver-based transport networks employing in-network caching (e.g. CDNs). Let us briefly describe how such systems work and summarize model assumptions and notation.

2 Content items (data) are split into packets, uniquely identified by a name, and permanently stored in one (or more) repository(ies). We consider a set of M different content items, of average size σ packets, grouped in K classes with the same popularity, i.e. content items of class k are requested with probability q k, k K. In the rest of the paper, we assume a Zipf popularity distribution, q k = c/k α, α >. Users retrieve a content item using a receiver-driven flow control protocol based on subsequent packet queries (denoted as Interests), triggering data packets delivery (see Fig.). Content requests for items in class k are assumed to arrive according to a Poisson process of intensity λ k = λq k (cfr. [9]). Considerations about ICP sensitivity to the Poisson assumption are reported in []. A name-based routing protocol guarantees that queries are properly routed towards a data repository, following one or multiple paths. Every intermediate node keeps track of outstanding queries in data structures called PITs (Pending Interest Tables), to deliver the requested data back to the receiver on the reverse path. In addition, nodes temporarily cache data packets in a LRU managed cache of size x packets (content store). Data may come from the repository, or from any intermediate cache along the path with a temporary copy of the data packet. Packets of the same content can therefore be retrieved in a multi-path fashion, i.e. from multiple locations with different round trip times, affecting the overall delivery performance. Independent of the considered topology, and for a given user, we denote P the set of paths relaying the user to content repository(ies). Each path is composed by intermediate hops i =,,... numbered starting from the user. Indeed, when an Interest is emitted and forwarded through a path, a copy of the requested packet can be found and retrieved at intermediate hops caches. We denote such sub-paths as routes and R φ identifies the set of routes of a given path φ. E.g. r φ i refers to the sub-path from the user to the i th hop on the path φ to the repository. The resulting rate and content delivery time are strongly affected by the notion of average distance between the user and the location of the requested content, which we will explicitly define through the virtual round trip time (VRTT), in analogy with the RTT of connection-based transmission protocols like TCP in Sec.V. A. Protocol Objectives III. ICP DESIGN A robust receiver-driven flow control protocol is responsible for realizing efficient data retrieval by controlling the receiver request rate and adapting it according to available network resources. In CCN, a transport session is realized by Interest queries progressively sent out by the receiver, routed by name towards a content store. Reliability: the first goal of ICP is to guarantee reliable data transfers by re-expressing Interests in case of Data packet losses. ICP schedules Interest retransmissions at the expiration of a timer τ, maintained at the receiver. In this way, the flow control protocol does not rely on losses as congestion notifications, but on delay measurements and timer expirations. Efficiency: the second objective of ICP is to minimize the completion time of data transfers. To this aim, ICP employs an AIMD (Additive Increase Multiplicative Decrease) mechanism for adapting the Interest window size as a mean to attain the maximum available rate allowed on the network path. Fairness: the third goal of ICP is to achieve fair bandwidth allocation among flows (namely, content retrievals), sharing the same route and, thus, the same limited down-link bandwidth. Let us specify the details of ICP design, driven by the objective to attain a globally efficient and fair equilibrium. B. Protocol description Data packets are identified by a unique name (the content name plus a segment identifier) and are requested via Interests in the order decided by the application. The receiver window, W, defines the maximum number of outstanding Interests a receiver is allowed to send. ) Interest Window Increase: W is increased by η/w at Data packet reception. This operation is implemented using integer arithmetic so that W is incremented by η (set to as default value) when a complete window of Interest is acknowledged by data reception. ) Interest Window Decrease: A congestion detection coincides with a timer expiration. The protocol reacts by multiplying W by a decrease factor β <, no more than once for a duration equal to the timer as in Fig.(b). Note that re-expression of an interest, after a time-out, is necessary to recover losses, even though data may just be delayed at the bottleneck. At the same time, retransmitted Interests sent just after a time-out can be filtered, if the PIT timer at the first node is not yet expired (see Fig. on the right). Fig.. Trellis diagram for window evolution: increase (left), decrease (right). ) Timers: Properly setting the Interest timer value τ and the PIT timer (which allows to remove a pending Interest at each node) is crucial in CCN. In general, τ must be trivially set larger than the minimum network delay, otherwise every sent Interest would trigger a timer expiration. In addition, a minimum τ value is necessary to guarantee high utilization of the available bandwidth, as precised by the analysis in Sec.IV- VI-A. A trade-off between a large τ value for efficiency reasons and a small τ value for faster window adaptation, is given by an adaptive timer based on the estimation of the minimum round trip delay, as common in TCP literature. Therefore, in the case of variable τ, ICP maintains round trip delay estimates at every data packet reception, by updating

3 RTT min and RTT max averaged over a history of samples ( Data packets in our implementation), excluding retransmitted packets. Each flow adapts its timer τ according to the rule: τ = RTT min + (RTT max RTT min )δ () with δ =.5 (as in [], where stability of the RTT estimation is also proven). Also, in our setting, we use an initial value of τ = ms and RTT max = RTT min initially after the first measured roud-trip delay. The PIT timer is fundamental in CCN to limit the number of pending Interests, hence the table size. The larger this value the higher the number of filtered Interests, not forwarded upstream towards content repository. On the other hand, a small PIT timer implies a large number of unnecessary retransmissions, since delayed packets arriving after PIT time-outs are discarded. Throughout the paper, we set the PIT timer never smaller than τ. IV. SINGLE BOTTLENECK ANALYSIS Consider the case of a single bottleneck link of finite downlink capacity C pkt/s relaying a user to a content repository. We assume for the analysis constant round-trip propagation delay R and Interest retransmission timer τ > R. We focus on a single content retrieval before extending the analysis to the case of n parallel content transfers. Interests are sent using the protocol described in Sec.III, with no maximum window limitations. We analyze the system evolution via a deterministic fluid model of the instantaneous rate at the receiver X(t) and the instantaneous queue occupancy Q(t) at the repository, both modeled as continuous time variables (see [] for an overview). Data packets are queued at repository output interface, as a result of down-link bandwidth limitations. The queuing delay is Q(t)/C, where Q(t) is the instantaneous queue occupancy in number of packets at time t. System dynamics are described by the following ODEs: dx(t) η = dt (R + Q(t)/C) βx(t) {R+Q(t)/C=τ}, dq(t) = X(t) C {Q(t)>} () dt The Interest rate is assumed to be linearly proportional to the Interest window W (t), that is X(t) = W (t)/(r + Q(t)/C) and hence it linearly grows proportional to the inverse of the square of the round trip time. A time-out triggers a window (hence, a rate) drop proportional to the decrease factor β <. We remark that the model defines a time-out at the instant when the round trip time is equal to τ to avoid unrealistic repeated window decreases. This models the fact that the protocol triggers no more than one window decrease over a time window of τ, after a timer expiration. Note that this fluid representation via ODEs has been extensively adopted in TCP modeling literature (e.g. []). Proposition.: The steady state solution of eqq.() is a periodical limit cycle of duration T, with < T < Cβτ ( β)η, and time average X = C, Q L Q C(τ R) where Q L = C(τ R) η ( C(τ R)β β ) and assuming that τ > τ min, with τ min = R + η T C ( β β ). The time average is defined as f = T f(t)dt. Proof: reported in [] due to lack of space. Observation.: From the steady state analysis we can draw the following conclusions: ) The value of τ must be chosen larger than the network delay R, otherwise every Interest request expires and the window decreases to the minimum value. ) At the same time, τ should be chosen large enough to accommodate more than the bandwidth delay product in order to fully utilize the link capacity. Here we focus on a large buffer regime allowing full bandwidth utilization. In case of small buffers, the queue may become empty, and the protocol would not assure full utilization. We neglect this second regime that can be avoided by correct parameter setting of the protocol. Window size [pkts] Queuing delay [ms] Time [s] Number of flows w (t) w (t) w (t) Q (t)/c Fig.. Interest window and queuing delay evolution for three data transfers, with R= µs, C= Mbps, τ = ms. Prop.. can be generalized to the case of multiple flows sharing the same link. Proposition.: Given a variable number of flows sharing the same link, generated according to a Poisson process N of rate λ, each one composed of σ packets on average, the expected flow rate in steady state is E[ X] = C( ρ), where ρ = λσ/c. Proof: reported in [] due to lack of space. To support analytical findings and to assess model accuracy, we developed a CCN packet-level simulator whose details are reported in []. In Fig. we show the simulated Interest window size evolution of three content retrievals starting at different instants, with the first transfer ending at t = 9 s. The queuing delay is also reported. Let us first observe the periodical behavior of window size and queue occupancy, whose limit cycle satisfies analytical prediction. Moreover, as expected, the frequency of queue occupancy oscillations is proportional to the number of parallel transfers n, since the queue input rate linearly increases with slope nη. Let us extend this analysis to the network case where multiple bottlenecks can exist and the distance between the user and the closest content replica is impacted by the caches along the route. To this aim, the next section introduces the CCN storage sharing model we introduced in [9]- []. Number of flows

4 V. DATA STORAGE SHARING MODEL As previously observed, each path is associated a metric that we denote as virtual round trip time that accounts for the average distance (in time) at which a Data packet is retrieved. Definition 5.: VRTT k represents the average time interval between the dispatch of an Interest and the Data packet reception for packets of content items in class k. This variable plays a similar role to round trip time for TCP connections in IP networks. When the time variable is not specified, VRTT k denotes the stationary value. VRTT φ k is defined for a given path φ from the user to a repository as a weighted sum of the round trip delays R(r φ i ) associated to the sub-paths r φ i, i.e. to the route between the user and the i th hop, with i =...N, where the weights correspond to the stationary hit probabilities, p k (r φ i ), to find a chunk of class k at hop i on the considered path, given that a miss was originated by all previous hops. For k =,..., K, VRTT φ k = i R(r φ i )( p k(r φ i )) p k (r φ j ) () i:r φ i Rφ j= To ease the notation in the case of a single path, we will replace r φ i by i, and VRTT φ k by VRTT k, whenever a single path is considered. Under these assumptions, in [9], we derived the miss probabilities in steady state for the a linear and a tree topology with no uplink bandwidth limitation. Note that a single path relies the user to the repository in these topologies, though the same analytical arguments allow to extend the hit probability computation to more complex topologies with multiple paths. We report here the main results of [9], generalized in [] to the case of possibly different cache sizes. Recall that for a given path φ, hop index i also identifies the route r φ i. Proposition 5.: Under previous assumptions, for a given path relaying a user to content repository, the stationary miss probability for packets of class k requested at the first hop, p k (), is given by p k () e λ m q kgx α () for large x, where /g = λcσ α m α Γ ( α) α, qk = c/k α, and x denotes the cache size at the first hop. As shown in [9], the miss process at node, and more generally i, can be well approximated with a Markov Modulated Rate process (MMRP) (i.e. a Poisson process at content level) of intensity µ(i) = K k= p k(i)λ k (i). For a linear topology the miss rate at node i also coincides with the input rate at node i+, i, i.e. µ(i) = K k= p k(i)λ k (i) λ(i + ), whilst for the symmetric binary tree topology λ(i + ) = µ(i), for i. The content popularity distribution also changes at hops i > because only missing packet requests are forwarded upstream, then for k =,..., K, and using q k q k (), i j= p l(j)q l. q k (i) = i j= p k(j)q k / K l= For a linear topology, the stationary miss probability values at hops i > are consequently derived. Proposition 5.: Given a linear topology with N hops between a user and the content repository a MMRP content request process with intensity λ and a popularity distribution q k, k =,..., K, < i N it holds i ( ) α xj+ log p k (i) = p k (j) log p k () (5) j= x j Notice that the same results hold for a hierarchical tree-like network topology. Proof: reported in [9]- []. VI. MULTI-BOTTLENECK ANALYSIS The virtual round trip time weights the delivery time associated to each route by the probability to retrieve a data packet from the given route. As a result of the interaction among flows, and of the transport protocol reaction triggered by time-outs, route delivery time varies over time on a short timescale. On the contrary, the probability to retrieve a Data packet from a given route, changes on a longer timescale. In fact, it depends on cache dynamics at different network levels, less sensitive to traffic variations. Such timescale separation is confirmed by our simulations and motivated by the stationarity of the global request process which impacts the hit/miss probabilities at different caches. Therefore, we can refer to the storage sharing analysis in Sec.V as a macroscopic model of storage sharing of CCN network dynamics, which gives us the hit/miss probabilities based on a stationary request process. On top of it, we provide, in this section, a microscopic study of the packet-level dynamics of content retrieval (flows), subject to the AIMD Interest rate control described in Sec.III. The analysis proves that ICP realizes the efficient and fair bandwidth sharing equilibrium described in [9]- []. Let us state the main results about rate and queue occupancies evolution (Sec.VI-A) for a route composed of N caches towards the repository. A. Main results Let us focus on the case of a single path between the user and the content repository, hence we omit the path index φ. We denote by n i,k the number of ongoing flows of class k over route r i and by n i K k= n i,k the flows sharing route r i. We generalize the single bottleneck analysis in Sec.II to the case of a path composed of N links, fed by content requests at the leaf node. We assume a negligible propagation delay with respect to the queuing delay at each link. The instantaneous rate X k (t) is modeled by a weighted sum of the rates achieved on different routes, X(i, t), corresponding to the retrieval of packets from different hops. We write the following set of state equations: dx(i, t) η = dt VRTT(i, t) βx(i, t) {VRTT(i,t)=τ}, dq i (t) = n i X(i, t) C i {Qi(t)>} dt N ( l ) + {Qj(t)=}n j X(j, t) + {Ql (t)>}c l l=i+ j=i+ where VRTT(i, t) = i j= Q j(t) C j, i =,..., N denotes the round trip time, given by the aggregated queuing delay associated to route r i, i.e. all links from the receiver to ()

5 the i th node. The average rate per class is then X k (t) = N i= X(i, t)( p k(i)) i j= p k(j). Note that the ODEs for X(i, t) are the same as in the single link case, with X k (t) = X(, t). Observation.: It is worth observing that content popularity affects the offered traffic on each link, as a consequence of request rate and hit/miss probabilities at different hops. This leads to a different rate X k (t). However, the rate that each flow gets on a given route, X(i), is the same for all popularity classes given that the delay associated to a route r i, VRTT(i, t), is the same for all k =,...K. Therefore, flows fairly compete on a given route for sharing bandwidth. The link load makes the number of flow in progress increasing and the fair share decreasing. Proposition.: The steady state solution of eqq.() is periodical with limit cycle of length T and it satisfies X k = N i ( p k (i)) p k (j) X(i), X(i) = γ(i) (7) i= j= where γ(i) is the max-min fair rate at the route r i. Proof: reported in [] due to lack of space. In contrast to TCP, ICP faces typically multiple bottlenecks for the majority of file transfers. We make this clearer through the following example. Example.: Consider the case of a single path composed by three caches relaying the user to the repository. Given that the number of ongoing flows on route i is n i, i =,,, it holds (7) where γ(i) vary according to C i /n i values. Four cases can be distinguished: - Case I: C /n < C /n, C /n γ(i) = C /(n + n + n ), i. - Case II: C /n < C /n, C /n, γ() = (C C )/n, γ(i) = C /(n + n ), i =,. - Case III: C /n < C /n < C /n, γ(i) = (C i C i+ )/n i, i =,, γ() = C /n. - Case IV: C /n < C /n < C /n, γ(i) = (C C )/(n + n ), i =,, γ() = C /n. Notice that, we solved eqq. to compute the fair rates on each route along the path, that depend on link capacities and the number of flows in progress on every route. VII. DESIGN GUIDELINES AND PROTOCOL ASSESSMENT In Sec.VI we have shown that, under proper parameter setting, the equilibrium achieved by the system, in terms of average rates and queue occupancies, is fully efficient and max-min fair as described in []. The choice of the parameters can be critical and can potentially lead the system to an inefficient regime, as already observed in Sec.IV. The present section provides guidelines for tuning protocol parameters, i.e. η, β and τ, and compare the protocol performance analysis of previous Sec. IV and VI to simulations on some simple scenarios. To this aim, we developed an event driven CCN packet simulator implementing: name-based (multi-path) routing, forwarding with output link queuing, in-network storage as well as ICP at the receivers. A. Parameters setting The additive increase parameter η, is set by default to, corresponding to an interest window increase of one packet per round-trip time. By increasing the window size more aggressively (η > ), a flow can achieve higher throughput, if and only if bandwidth resources are not fully utilized, as observed in [5]. However when bandwidth is efficiently utilized, we show in Sec.VI, that η affects only the limit cycle duration and not the average rate. Similar considerations hold for the multiplicative decrease parameter β, set by default to.5, such that the Interest window is halved upon the expiration of the timer. As for η, a smoother window decrease, setting β.5, results in higher flow throughput as long as bandwidth is not fully utilized. As reported in Sec.VI, in case of efficient bandwidth sharing, β affects the limit cycle duration and not the time average. The most critical parameter is the Interest retransmission timer τ, that must be chosen not smaller than the minimum network delay to have non zero throughput, and large enough to queue at least the bandwidth delay product along the path, to have % utilization. Moreover, an arbitrarily large τ would let ICP converge very slowly, which motivates the choice to adaptively set τ according to eq.(). B. Numerical results In this section we asses protocol performance in some simple scenarios, comparing simulations to the optimal working point calculated in Sec.VI. Constant τ: We consider a linear network with tree links in a multi-bottleneck configuration, assuming negligible propagation delays; σ = 5, packets (packet = kb ) and cache sizes of 5, packets. Content items of class k are requested according to a Poisson process of intensity λ k = q k content/s, with q k = c m k, where m = M/K = 5 is the α number of content items in the same class, and α =.7 the Zipf parameter. Fig.(a) reports the average delivery time (top) and virtual round trip time, VRTT k (bottom), measured for content retrievals in the ten most popular classes, k =,...,, for different values of τ. As predicted by the model, VRTT k is shown to be proportional to τ in Fig.(a) (bottom). The measured delivery time is compared to the optimal reference curve, indicating full utilization and fair allocation of resources; moreover it is shown to be negatively affected by a too small τ value, especially for classes k >, since longer routes r, r (of two/three hops) does not fully utilize the bottleneck capacity C. Conversely, most popular data transfers (k = ) get slightly better performance than what predicted, as the τ value is well set for the shorter route r they mainly utilize. We argue that each data transfer would benefit from adapting τ to the measured delay from used routes. Variable τ: In Fig.(b) we present a set of simulations, where each receiver estimates the value of τ, by implementing the algorithm described in Sec.III-B. Globally, the queuing delay converges to a much smaller value than that obtained using a constant τ, as a consequence of the fact that τ k, converges to value that is slightly larger than VRTT k, function of the popularity class and, hence, of the data distribution along the path.

6 VRTT k [ms] τ = ms τ = ms τ = 8 ms τ = ms Model Class k (a) VRTT k [ms] Case I Case II Case III Simulation Model Class k (b) φ users (case II) φ users (case III) Multi-homed users Class k Simulation Model Buffer size [pkts] (c) Fig.. (a) Delivery time and VRTT k in case II with C = C = Mbps, C = Mbps with fixed τ, (b) with adaptive τ case I: C = 5Mbps, C = C = Mbps; case II: C = C = Mbps, C = Mbps; case III: C = C = Mbps, C = 5Mbps (c) Multi-Homed network scenario. ρ =.8 ρ =.9 Multi-homed scenario: As already described in Sec.II, routes are built following the reverse path of the Interests, routed by the name to the closest hitting content store. Every receiver can easily exploit multiple forwarding faces by load balancing Interests on the different faces. Let us consider a network topology with multiple repositories, where one group of users is multi-homed, namely it uses two disjoint paths, φ and φ, with C φ = C φ = Mbps, C φ = Mbps and C φ = C φ = Mbps, C φ = M bps. Multi-homed users share bandwidth resources with other single path users exploiting φ or φ only. Results are reported in Fig.(c) (top). A good agreement can be observed between simulation results and analytical predictions both in the single and multi-path case. Moreover, as expected, VRTT k and delivery time experienced by multi-homed users, are significantly reduced w.r.t. those perceived by φ /φ single path users as a consequence of the higher throughput achieved by using both path in parallel. Buffer Sizing: Throughout the paper we have considered networks with unlimited buffer capacity. System dynamics in a small buffer regime have not been analyzed due to lack of space, for completeness, we report in Fig.(c) (bottom) the average delivery time measured in a single bottleneck scenario at two loads,.8,.9, under different buffer sizes. The loss of performance due to little buffering can be significant. However, we believe that there is no clear advantage in saving output buffer memory in CCN, where technological limitations are mainly faced in the content store []. VIII. CONCLUSIONS AND OPEN ISSUES In this paper we have designed ICP, an interest control protocol for content-centric networks driven by a windowbased AIMD controller of the Interest rate expressed at the receiver. ICP is conceived to control Data packet retrieval from the different routes (subpaths) that constitute each path from a user to a content repository. An analytical characterization of user s stationary rate and queues dynamics is provided, under general traffic demand, allowing to understand the impact of CCN transport and caching dynamics on user performance. The key takeaway of our analysis is that the present ICP design realizes optimal statistical bandwidth sharing guaranteeing efficient and fair bandwidth utilization ( []). Guidelines are also provided to optimally set protocol parameters. Protocol performance are evaluated by packet-level simulations, corroborating analytical results on a hierarchical network topology. We expect future experimentation at a larger scale to confirm such good premises. Finally, we plan to extend multi-path analysis beyond the case of independently controlled paths to better exploit the intrinsic potential of multi-path communication in CCN. ACKNOWLEDGEMENTS This work has been partially funded by the French national research agency (ANR), CONNECT project, under grant number ANR--VERS- ṘEFERENCES [] V. Jacobson, D. Smetters, J. Thornton, M. Plass, N. Briggs, and R. Braynard, Networking named content, in Proc. of ACM CoNEXT, 9. [] T. Koponen, M. Chawla, B. Chun, A. Ermolinskiy, K. Kim, S. Shenker, and I. Stoica, A data-oriented (and beyond) network architecture, in Proc. of ACM SIGCOMM, 7. [] Information centric network projects, 5 Related-Projects. [] B. Ahlgren and al., Design considerations for a network of information, in Proc. of ACM CoNEXT, 8. [5] P. Jokela, A. Zahemszky, S. Arianfar, P. Nikander, and C. Esteve, LIPSIN: Line speed Publish/Subscribe Inter-Networking, in Proc. of ACM SIGCOMM, 9. [] S. Arianfar, P. Nikander, and J. Ott, On content-centric router design and implications, in Proc. of ACM ReArch,. [7] U. Lee, I. Rimac, and V. Hilt, Greening the internet with content-centric networking, in Proc. of ACM E-Energy,. [8] G. Carofiglio, V. Gehlen, and D. Perino, Experimental evaluation of storage management in content-centric networking, in Proc. of IEEE ICC,. [9] G. Carofiglio, M. Gallo, L. Muscariello, and D. Perino, Modeling data transfer in content-centric networking, in Proc. of ITC,. [] G. Carofiglio, M. Gallo, and L. Muscariello, Bandwidth and storage sharing performance in information centric networking, in Proc. of ACM SIGCOMM Workshop on ICN,. [] L. Zhang and al., Named data networking (ndn) project,., http: //named-data.net/ndn-proj.pdf. [] G. Carofiglio, M. Gallo, and L. Muscariello, Icp technical report. [] A. Kuzmanovic and E. Knightly, Tcp-lp: low-priority service via endpoint congestion control, IEEE/ACM Trans. on Networking,. [] M. Ajmone Marsan, M. Garetto, P. Giaccone, E. Leonardi, E. Schiattarella, and A. Tarello, Using partial differential equations to model tcp mice and elephants in large ip networks, in IEEE INFOCOM,. [5] A. Kuzmanovic and E. W. Knightly, Receiver-centric congestion control with a misbehaving receiver: Vulnerabilities and end-point solutions, Elsevier Computer Networks, 7.