Research Progress for the Interest Control Protocol of Content-centric Networking

Transcription

1 Sensors & Transducers 2013 by IFSA Research Progress or the Interest Control Protocol o Content-centric Networing GAO QIANG, WANG WEIMING, ZHOU JINGJING Zhejiang Gongshang University o Institute o Networ and Communication Engineering, Hangzhou, , China gaoqiang@pop.zjgsu.edu.cn, wmwang@zjsu.edu.cn, zhoujingjing@mail.zjgsu.edu.cn Received: 28 April 2013 /Accepted: 19 July 2013 /Published: 31 July 2013 Abstract: Content-centric networing (CCN) brings a new change in today's Internet communication mode by addressing named-data instead o host locations. Another major eature is the possibility to added storage capabilities into the networ. Now, the ocus o this wor is on the design o a receiver-driven Interest control protocol or CCN. ICP realizes a window-based Interest low control, achieving ull eiciency and airness under proper parameters setting. In order to research ICP in-depth, in this paper, we summary the recent research progress o it. ICP is divided into several parts: content request process, bandwidth sharing model, content request aggregation, storage sharing model. We mae these process modeling, research and analysis in linear and binary tree networ topology separately, and conirm the accuracy o the models and its characteristics through the numerical analysis, inally give the research trend or ICP. Copyright 2013 IFSA. Keywords: Content centric networ (CCN), Interest control protocol (ICP). 1. Introduction The design principle and structure o today's Internet originated in the in the 1960s and 70s. The goal o networ design is to solve the shared hardware resources. Thereore, the main purpose o the communication is to connect the two hosts, and the need to determine the speciic location o a device. However, with the rapid development o inormation technology, computer hardware costs a signiicant reduction in shared hardware resources demand have aded, and the subject o networ applications has turned to text messages and video content, and content services is slowly becoming a main networ service body. The content o the Internet service has been more attention. In content services networ, people do not care about what computer to provide content inormation, and is only concerned about the speed o access to the content as well as the reliability and security o the content. So, content service has become the main networ services. Under this premise, people do not care which computer provides the content inormation, but only concern about the speed o access to the content as well as the reliability and security o the content. However, the currently widely used in the Internet's TCP / IP is still a Host-to-Host communication model, this model is obvious shortcoming to post and access to inormation-based Internet. In this case, the point-to-point (p2p) technology and content delivery networ (CDN) technology, to some extent, ease the user's content / inormation sharing needs. However, any content delivery mechanism in the current networing is not completely overcome the deects o the underlying mechanism, resulting in a waste o resources, and requires a complex mapping o the content and "position". Article number P_

2 In order to adapt to this change in demand, new networ architecture should be established. Thereore, the Inormation Center Networ (ICN, Inormation Centric Networing) is proposed, and its goal is to solve all the problems existing in the internet today. It has been proved that ICN is better able to solve the various problems in today's Internet. In these uture networs research projects, Palo Alto research Center s (PARC) Content-centric networing (CCN) [1] has more advantages. Currently, many research institutions are involved in the content-centric networ architecture, Such as UCLA, and Bell Labs [7], Europe CONNECT project team [2] and so on. The rest o this article is organized as ollows. In Section 2 describes the CCN architecture, the dierence between CCN and the TCP / IP protocol. In Section 3 describes the research progress o the ICP protocol requests rom the research ramewor and content transmission bandwidth sharing, shared storage and iltering o the routing nodes several processes to describe. In Section 4 is a summary o the thesis. 2. CCN System Introduction The basic meaning o the Content-centric networing is that the demand o the entire networ is content, not the host. It undamentally changed the IP pacet s encapsulation structure and addressing mode, the pacet header no longer identiies the address, but the content name. CCN communication is driven by the consumers o data. There are two CCN pacet types [4], Interest and Data (Fig. 1). Fig. 1. CCN pacet types. The entire inormation transmission process is that the request pacet deines the title o the contents contained in a pacet and the pacet is transmitted to all directions, then the device o an adjacent node which can provide the requested content will sent data to the requesting node through the content data pacage. It can be seen that the CCN entire process is no longer concerned about the location, but quicly get the most rom the neighboring nodes in a minimum price. So, CCN provides a new protocol stac (Fig. 2). Fig. 2 compares the IP and CCN protocol stacs. CCN architecture shape and TCP / IP networ is very similar, the lower layer protocols are designed or adaptation o the underlying physical lins and communication, the upper layer protocols are designed to correspond to the related application. The main dierence is that the CCN using the Content chuns instead o the IP and using named data instead the naming o physical entities. IP protocol's success lies in its relaxed requirements or second layer protocol, CCN inherited it, and CCN can be implemented in any o the underlying protocols, can even architecture above the IP layer. In addition, the content-centric networ routing node built storage unction used to cache data pacets which can speed up access to other users and reduce networ traic. Fig. 2. The compare between TCP/IP and CCN. The ICP protocol [5] o CCN describes the entire inormation communication process: air sharing o the transmission, and the routing node storage and iltration process rom the transmission o the request pacet, the bandwidth on the lin. We will describe the progress o the current study as ollows. 3. ICP Introduction The CCN s ICP protocol [5] describes the entire inormation communication process. The client sends a request pacet, transmitted via the lin, reaches a routing node, the routing node o the query sequence are sequentially processed (the detailed procedure as ollowed), then sends to uplin by a node (in CCN, only the request pacets are routed), until retrievals needed data pacets, then the pacet request reverse transmission path o the pacet routing, service receiving end. The ICP process is divided into several parts: content request process, lin bandwidth sharing transmission process, storage and iltration process o the routing nodes. For the beneit o research, we model these sub-processes, and discuss the ICP s availability, eiciency and airness based on linear and binary tree networ topologies Node Model Analysis Interest Pacet Routing Analysis From Fig. 3, the typical CCN node model includes the contents o memory (CS), pending requests table (PIT) and orward orwarding (FIB). 88

3 CS is similar to IP routers cache, but uses dierent caching strategies. Because each o the IP pacets belong to a single point-to-point session, no value to the other session, so it uses the MRU replacement strategy which empty the inormation stored ater the completion o a session. CCN each pacet is selidentity and sel-validation, so there is also value or other users, CCN will uses the LRU or LFU replacement policy to maximize store important inormation, in order to reduce the content download delay and networ bandwidth. PIT records the uplin routing path o request pacets, so the content data pacage can be bactrac spread bac to the requesting node according to PIT recorded inormation trac. When the content requested arrives at the receiving node, PIT will delete this entry. FIB and the IP router s processing mechanism are similar, but it can be orwarded to the plurality o directions at the same time request. When an Interest pacet arrives on some ace, a longest-match looup is done on its Content Name. The index structure used or looup is ordered so that a Content Store match will be preerred over a PIT match which will be preerred over a FIB match. The speciic operation is as ollows: Fig. 3. CCN orwarding engine model. I CS contains a request pacet, it will directly sends the corresponding content to the requested port, and discards the request pacet, otherwise it will continue to query the PIT. Otherwise, i PIT contains entries associated with the content name, it will let request port added to the request port list, and discards the request pacet, otherwise it will continue to query the FIB. Otherwise, i FIB contains entries associated with content name, it will orward the pacet to the next CCN node in accordance with the instructions o the FIB. I the rest o the port is not empty, then it will orward request to all the rest, and ormat new list o entries and ports in the FIB. I there is no match or the Interest it is discarded (this node does not have any matching data and does not now how to ind any) Data Pacet Routing Analysis Data pacet processing is relatively simple since Data does not require routing, ollowed by only PIT port records the request data pacet transmission trajectory to reach the initial client. When a pacet arrives routing nodes, the longest match its content name query. Described in the PIT match the pacet ater the request when the CS match the pacet duplication, to remove it; PIT port o the FIB matching the node does not match, the pacet is unsolicited delete it; pacet orwarding path, request the end through the node to accept Assumptions and Notation For the various parts o the ICP protocol, we reerence [5, 7, 8] and summarize all o the models. The main assumptions are the ollowing: 1. M dierent content items are equally partitioned in K classes o popularity, i.e. is requested with probability q, 1. We assume a Zip popularity distribution, q c K, c ( 1 1/ ), with Content items are divided into chuns (P size) and the content size distribution is geometric with average σ chuns. 89

4 3. N networ nodes (or levels or tree topologies). Node i is equipped with a cache o size xi chuns. 4. The request arrival process is modeled as a Marov Modulated Rate Process (MMRP) [3] and the process is according to a Poisson process o intensity q 5. We deine virtual round trip time o class, VRTT, is similar to the round trip time or TCP connections in IP networs. It describes the average time o a chun to send the request and is accepted. than once or a duration equal to the timer as in Fig. 4 (b). Interest timer value: τ should be set larger than the minimum networ delay, and a minimum τ value is necessary to guarantee high utilization o the available bandwidth. In the case o variable τ, ICP maintains round trip delay estimates at every data pacet reception. Each low adapts its timer τ according to RTT ( RTT RTT ), 0.5 (2) min max min N i1 VRTT R (1 p ( i)) p ( j), 1,..., K i i1 j1 (1) VRTT is deined as a weighted sum o the round trip delays Ri associated to node i, and the weights are related to the stationary hit probabilities 1 p ( i) Content Request Process Notation Fig. 4. Linear (a) and binary tree (b) topologies with bandwidth-limited down-lin Content Request Aggregation The content request process is structured in two levels, content and chun. The process is modeled as a Marov Modulated Rate Process (MMRP): the request or content items is according to a Poisson process o intensity: q. A request or content is exactly the content o the irst chun, and when a chun is received, the other new chun is going to sent until the contents o the last chun. ICP realizes a window-based Interest low control, and it uses the AIMD (Additive Increase Multiplicative Decrease) mechanism to adapt the Interest window size to attain the maximum available rate allowed on the networ path. Interest window increase: W is increased by η/w at Data pacet reception. Interest window decrease: The protocol reacts by multiplying W by a decrease actor β< 1, no more CCN s a basic eature o blocing requests lood is the aggregation, which it can prevent the transmitted request or the same data again by the node o the PIT records pending requests chun trajectory. PIT time window Δ: the value is used to limit the number o pacets in its pending request. The larger o the value, the more the number o ilter out a request pacet; smaller the value, means that there are more unnecessary requests continue to be routed. The request aggregation clearly impacts the miss rate o a given cache. In act, when a request chun arrives at a node, i the corresponding chun is stored in the node, it will generate a hit, otherwise it generates a lost. For the latter, the request is iltered only i a previous request or the same chun has been emitted and the chun has not been received yet. Given the request process, the iltering probability associated to class at the irst hop is, 1 b pilt, (1) 1,..., K 1 (11/ ) b b e The proo is reported in [11]. (3) 90

5 3.5. Bandwidth Sharing Model In CCN, a receiver can download data rom multiple nodes and share bandwidth with other concurrent transers. User-perceived perormance depends on the parallel download bandwidth sharing. A generally accepted airness objective is max-min, which can achieve rate equalization. Every route i is characterized by a number o parallel transers, N i, sharing the same route and thereore the limited bandwidth. N i is a random process that varies according to the demand λ and the service rate each transer in progress gets. Ni is a birth and death Marov process, birth rate: ( i1) ( i), ( (0) ) death rate: n ()/( i ) i ( i1) ( i) denotes the rate o the Poisson process o content requests satisied by node i, under the assumption o a MMRP miss process o intensity () i at node i. The death rate, ni ()/( i ), is determined by the max-min air shared bandwidth () i in bps associated to each o the ni content transers in parallel over route i. orwarding tables are implemented by Carzaniga et al. in content-based networing simulator [6]. For the case o a single node, we present the experimental case is that a population o M = content items, and K = 400 classes, each one with m = 50 items. Content items are split in chuns o 10 B each. Users generate content requests according to a Poisson process o intensity λ= 40 content/sec, and the interest transmission window size is W = 1. The cache o size x = chuns (2 GBs) which implements the LRU replacement policy. Each cache is assumed to be initially empty, while statistics are collected in steady state only Storage Sharing Model The research about storage sharing model is based on above three models: content request process, content request aggregation and bandwidth sharing model, so we validate numerical analysis through the storage sharing model. Except 3.2 assumptions, we give the others as ollows: 1. K popularity classes with the same number o content items m = M/K in each one. 2. Non-unitary content size (σ) Storage Model Analysis at Node 1 We irst analyze the irst node which means the irst rom the content request point. Given a MMRP request arrival process with intensity λ, and average content size σ, be x > 0the cache size in number o chuns, then the stationary miss probability or chuns o class is, 1/ g c m (1 ) (4) 1 1 We use an ad-hoc C++ event-driven simulator[8] to conirm the accuracy o the model and assume a simple transport protocol which it uses a ixed window size or chun requests as in CCN. Nodes Fig. 5. Miss probability as a unction o class popularity (a); miss rate with and without request iltering or α= 2 (b) [8]. In Fig. 5 (a) [8] we can see that the greater the value o α, the miss rate is lower, and or high popular level (1 is the highest) o the content, it s miss rate is low, while the lower popular Level o the content, it s miss rate is getting higher and higher, almost, that is certainly lost. Fig 5 (b) [8] shows the miss rate with request iltering is lower than without. Through these two simulation graphics, we can conirm the accuracy o the eq.4 or predicting the probability o loss. 91

6 Storage Model Analysis at Networ Topologies In this section we consider the topologies reported Fig. 4, in order to derive the stationary miss probabilities p () i at hop i > 1, we need the ollowing auxiliary result. Given a MMRP request arrival process with rate λ(i) and popularity distribution: q () i i1 p ( j) q / K l1 ( ), 1,... p j q K j1 l1 j1 l l and deined Si(0, t) the number o dierent chuns th requested in the open interval (0, t] at the i node. Be μ(i) the miss rate at node i () i () i p () i q () i, then it holds [ S (0, t)] () i 1 t ( i1) (5) 1/ gi ( ) lim i 1 c m (1 ) t i.e. gi ()/ g ( i 1)/ () i, being g(1) g.the proo is reported in [9] Miss Rate Characterization without Request Aggregation Given a cascade o N caches as in Fig. 4 (a) and a MMRP request arrival process, then 1 < i N it holds gi () log ( ) ( ) log (1) p i i1 p l g(1) l1 p i1 p()log l p(1) l1 (6) Similarly, one can study the binary tree topology in Fig. 4 (b), that has no dierence with the line topology in the homogeneous case expect that λ(i)=2μ(i-1). Given a homogeneous binary tree with 2 N 1caches (that is with N levels) as in Fig. 4(b) and a MMRP requests arrival process, then 1 < i N it holds p () i p (1) p () i i1 p () 1 1 l p ilt, (1) l (8) For the topology in Fig. 4 (b) the request aggregation can tae place at several hops depending on traic and cache parameters. Given a binary tree with 2 N 1 caches (that is with N levels) as in Fig. 4 (b) and a MMRP requests arrival process and an aggregation timescale, then it holds 1 b ( i) p ilt, () i, 1,..., N 1 (11/ ) b ( i) (9) with / m 2 ( i 1)/ m b (1) e, b ( i) e, i 1. (1)(1, (1)) 1 p pilt i i () i 2 ( i1) p ()(1 i pilt, ()) i i i 1 (10) The proo is provided in [9]. For the case o topology, we present the experimental case is that N = 3 levels binary tree, where data is stored at the root o the tree. Each lin has the same capacity o 10 Gbps and the same round trip delay equal to 2 ms. The Zip parameter α= 2, and chun size is 10 B. Other parameters are same to the above case. Fig. 6 compares the miss probabilities at dierent nodes (rom the 1st to the 3rd level) without request aggregation. The comparison summarizes the good match between model predictions Eq.(4) at level i = 1, Eq.(6) at level i > 1. we can be see that the requests or content items o the most popular class are almost completely served by caches o irst nodes, so the miss probability or class = 1 is nearly 1 at i>1. Content items o class 2 are mainly cached at irst and second level, whereas less popular classes, represented in the queue o the curves, are very rarely cached as hardly requested. i1 () igi () log p ( i) p ( l)log p (1) i1 l1 ( i1) g(1) l1 p ()log l p (1) (7) Miss Rate Characterization with Request Aggregation Given a cascade o N caches as in Fig.4 (a), a MMRP request arrival process and an aggregation timescale, then it holds Fig. 6. Miss probability at dierent levels or binary tree topology, with no iltering [8]. 92

7 4. Conclusions Content-centric networing is the research ocus o the next generation networ. In this paper, we summary the recent research progress o ICP. We mae these process modeling, research and analysis in linear and binary tree networ topology separately, and conirm the accuracy o the models and its characteristics through the numerical analysis, inally give the research trend or ICP. Through the research, we ind that ICP protocol s a signiicant eature is the storage unctions embedded in the networ, it can help reduce content download latency and networ bandwidth usage greatly improve bandwidth utilization perormance. Another is that CCN transport provides Disruption Tolerant Networing. The multipoint nature o data retrieval by Interest provides lexibility to maintain communication in highly dynamic environments. Any node with access to multiple networs can serve as a content router between them. These characteristics do provide eective help to resolve the current Internet problems. But, CCN technology is still in its beginning stages. So, the ICP protocol is not very mature, and the protocol s process and topology are relatively simple, and the protocol traic control problem also needs more in-depth study. The next step o the research is continues to improve the ICP protocol, and over a period o time it will also be the hot issues o the next generation networ research. Acnowledgements This wor has been partially unded by National "973" project, under grant number 2012CB315902, and by NSFC (Natural Science Foundation o China), under grant number , and by Youth Science Fund o NSFC, under grant number , and by Natural Science Foundation o Zhejiang, under grant number Y , and by Public Interest Research social development projects, under grant number 2012C Reerences [1]. Project CCNxTM. [2]. CONNECT [3]. T. E. Stern and A. I. Elwalid, Analysis o separable Marov-modulated rate models or inormationhandling systems, Advances in Applied Probability, Vol. 23, No. 1, 1991, pp [4]. Van Jacobson, Diana K. Smetters, James D. Thornton, Michael F. Plass, Nicholas H. Briggs, Rebecca L. Braynard, Networing Named Content, in Proceedings o the ACM CoNEXT, [5]. Giovanna Caroiglio, Massimo Gallo, Luca Muscariello, ICP: Design and Evaluation o an Interset Control Protocol or Content-Centric Networing. in Proceedings o the INFOCOM Worshop, [6]. A. Carzaniga, M. Rutherord, and A. Wol, A routing scheme or content-based networing, in Proceedings o the INFOCOM 04, [7]. Giovanna Caroiglio, Massimo Gallo, Luca Muscariello. Bandwidth and Storage Sharing Perormance in Inormation Centric Networing, in Proceedings o the ACM SIGCOMM Worshop (ICN 11), [8]. Giovanna Caroiglio, Massimo Gallo, Luca Muscariello, Diego Perino, Modeling data transer in content-centric networing, in Proceedings o the (ITC23), [9]. G. Caroiglio, M. Gallo, L. Muscariello, and D. Perino, Modeling data transer in content-centric networing (extended version), in Research report, available at /muscariello, Copyright, International Frequency Sensor Association (IFSA). All rights reserved. ( 93