Popularty Based Adapte Content Delery Scheme wth In-Network Cachng Jeong Yun Km, Gyu Myoung Lee, and Jun Kyun Cho To sole the ncreasng popularty of deo streamng serces oer the Internet, recent research acttes hae addressed the localty of content delery from a network edge by ntroducng a storage module nto a router. To employ n-network cachng and persstent request routng, ths paper ntroduces a hybrd content delery network (CDN) system combnng noel content routers n an underlay together wth a tradtonal CDN serer n an oerlay. Ths system frst selects the most sutable delery scheme (that s, multcast or broadcast) for the content n queston and then allocates an approprate number of channels based on a consderaton of the content s popularty. The proposed scheme ams to mnmze traffc olume and achee optmal delery cost, snce the most popular content s delered through broadcast channels and the least popular through multcast channels. The performance of the adapte scheme s clearly ealuated and compared aganst both the multcast and broadcast schemes n terms of the optmal n-network cachng sze and number of uncast channels n a content router to obsere the sgnfcant mpact of our proposed scheme. Keywords: Content delery network, n-network cachng, request routng, content popularty. Manuscrpt receed Sept. 23, 203; resed Feb. 7, 204; accepted June 27, 204. Ths research was supported by the ICT Standardzaton program of MISP (The Mnstry of Scence, ICT & Future Plannng). Jeong Yun Km (correspondng author, jykm@etr.re.kr) s wth the Communcatons & Internet Research Laboratory, ETRI, Daejeon, Rep. of Korea. Gyu Myoung Lee (G.H.Lee@ljmu.ac.uk) s wth the Lerpool John Moores Unersty, Lerpool, UK. Jun Kyun Cho (jkcho59@kast.edu) s wth the Korea Adanced Insttute of Scence and Technology (KAIST), Daejeon, Rep. of Korea. I. Introducton In a content delery network (CDN), a CDN serer s tradtonally used to reduce traffc on the Internet backbone by offloadng traffc requests from the orgn serer. Howeer, sttng outsde networks proded by Internet serce proders (ISPs), a CDN serer cannot reduce traffc on the transt or peerng lnks that connect the ISP network wth the Internet backbone and other ISP networks []. As demand for content access and delery oer the Internet ncreases, nnoate CDN archtectures and technologes are becomng ncreasngly mportant to effcently cache and dstrbute the surgng amount of deo content. To mnmze delery latency and nter-isp traffc, a lot of recent researches address localzed delery of large content olumes from a network edge by ntroducng a storage module nto network enttes (for example, a content router) [2] [3]. In other words, a content router can be allowed to prode nnetwork cachng and localzed delery whle contnung to support ts basc features such as packet forwardng and routng. Therefore, from the ewpont of the desgn of a content router, the optmal n-network cachng sze should be carefully determned to mnmze the performance degradaton that results from the ntroducton of such a storage module. In general, content delery schemes can be classfed nto three major types. Frst, a uncast scheme does not approprate well at a large scale and s, therefore, not dscussed further n ths paper. Second, a multcast scheme allows a number of requests for the same content to be grouped together and sered by a sngle multcast stream. In a batchng-based multcast scheme [4] for example, seeral content requests are delayed for a perod of tme before fnally serng the resultng batch a a multcast stream. In a patchng-based multcast ETRI Journal, Volume 36, Number 5, October 204 204 Jeong Yun Km et al. 89 http://dx.do.org/0.428/etrj.4.03.0090
scheme [5] [6], a content request s frst sered by a uncast stream and then joned back to a multcast stream. Thrd, a broadcast scheme [7] can broadcast content on dedcated channels at a pre-defned schedule. Owng to the lmtatons of content cachng and content delery capabltes, content routers seem ery unlkely to cache all content. Howeer, t would be better to cache and deler a prefx (that s, the begnnng porton of the content), f ts length s suffcently short. In addton, prefx cachng has a number of adantages, such as a reducton of both delery latency to clents and traffc olume oer networks [5], [8] [9], partcularly compared to the threshold-based multcast scheme runnng on a centralzed serer [6], [0]. Therefore, the CDN serer can only deler the suffx that s the remanng porton other than the prefx to multple clents through a sngle multcast stream. Our preous work n [] showed that the performance of a patchng-based multcast scheme s much better than that of batchng-based multcast schemes. Howeer, the former requres that content routers perform relately complex processng operatons. Ths s caused by the occurrence of changes n suffx lengths, whch s due to the araton n the arral tmes of suffx requests. Thus, compared to the latter scheme, whch has a fxed suffx length, patchng-based multcast schemes need more complex operatons. Based on ths context, ths paper manly focuses on patchng-based multcast and broadcast schemes. Proxy-asssted multcast schemes [5], whch combne proxy prefx cachng wth multcast schemes, such as batchng and patchng, are generally known as ther system control s smpler than that of broadcast schemes. Such schemes can collect more requests for the same content because they are sered by a sngle multcast stream. On the other hand, proxyasssted broadcast schemes [7] can sgnfcantly reduce the network resource requrements as well as serce latency by broadcastng content to dedcated multcast channels. Howeer, most research has focused on deelopng multcast schemes for generally mnmzng the aggregate network bandwdth rather than the network bandwdth consumed by only proxy serers. The request-routng system (RRS) used n a tradtonal CDN system s used to redrect clent requests to the closest surrogate by consderng network proxmty to prode fast delery [2] [3], [2] [3]. Ths paper frst presents detaled operatons of a persstent RRS that can redrect all clent requests for the same content to a partcular content router once the router s chosen from the frst request. Therefore, such requests can consume only a sngle multcast stream durng ther prefx lengths, thereby reducng the amount of network resources used. In addton, the persstent RRS can prode a fner granularty (for example, content chunk leel) than that of the orgnal RRS (for example, content fle leel). Wth the persstent RRS and n-network cachng, ths paper ntroduces a hybrd CDN system that combnes noel content routers n the underlay wth the CDN serer n the oerlay. In addton to ths, the hybrd CDN system s capable of prodng adapte content delery. As an effcent delery scheme s adaptely selected accordng to content popularty for the oerall performance gan, the proposed popularty-based content delery scheme can sgnfcantly reduce delery latency and traffc olume oer the network. Gen the number of multcast channels n the CDN serer, we address the problem of both mnmzng the aerage number of channels (the requred capacty) at the content routers and determnng the optmal prefx length (that s, n-network cachng sze). We also ealuate and compare the performance of the proposed popularty-based adapte scheme wth other content delery schemes to hghlght the fact that the proposed one clearly has performance mproement aganst both the multcast and broadcast schemes coupled wth n-network cachng. The remander of ths paper s organzed as follows. The CDN system model s brefly presented n Secton II. Secton III descrbes a popularty-based adapte content delery technque wth n-network cachng. In Secton IV, we ealuate the performance of content delery schemes under aryng condtons. The paper s concluded n Secton V. II. System Model We llustrate the hybrd CDN system, whch conssts of an orgn serer, a CDN serer, a persstent RRS, and content routers, n Fg.. A group of clents receng content delered across networks from the CDN serer through the content routers are consdered. The Hypertext Transfer Protocol (HTTP) s used for descrbng the requested content by ts unform resource dentfer (URI) [4]. In general, the orgn serer s managed by the content proder and located n the data center. It also stores content that s dstrbuted to both the CDN serer and content routers before such a request s made. Clent Clent Uncast channel In-network cachng nsde content router Selectng best content router Content router Content router RRS Multcast channel CDN serer Fg.. Hybrd CDN system archtecture. Dstrbutng content Orgn serer 820 Jeong Yun Km et al. ETRI Journal, Volume 36, Number 5, October 204 http://dx.do.org/0.428/etrj.4.03.0090
Clent (8) Suffx response Prefx request Prefx response Suffx request Clent Suffx response () Content request (2) Redrect content request RRS (3) Prefx request (6) Suffx request (5) Prefx response (4) Suffx request Content (7) Suffx response router Content router Suffx request Suffx response CDN serer Orgn serer Fg. 2. Multcast delery scheme wth n-network cachng and operaton. () Content request (2) Redrect content request RRS (3) Prefx request (6) Suffx jon (5) Prefx response (4) Suffx request Content router Clent (7) Suffx response Suffx jon Prefx request Prefx response Suffx request Content router Clent Suffx response Suffx broadcast Suffx broadcast CDN serer Orgn serer Fg. 3. Broadcast delery scheme wth n-network cachng and operaton. Thus, the ISP s aware of what both the CDN serer and content routers hae cached [], [5]. The CDN serer can deler the requested suffxes to the clents through multcast channels. In addton, the content router s a network element that acts as a regular router. It can also cache and deler the nnetwork cachng prefx to the clent, though wth a buffer of lmted sze, through uncast channels. In Fgs. 2 and 3, the persstent RRS s used to locate the best content router, for a partcular clent, whle prodng the granularty of the content chunk leel n step. If the request s satsfed, then the RRS can return (n step 2) a status code, such as HTTP 300 Multple Choces, n ts response to nform the clent of the new URIs of both the content routers and the CDN serer. Such URIs also ndcate the content name and ts range namely, the content chunk. Ths paper fundamentally assumes that content can be dded nto two parts: a prefx and a suffx. The clent should then smultaneously ressue ts prefx and suffx requests wth two or more HTTP GETs to the content routers and CDN serer, respectely. If both can return the requested chunk (that s, prefx and suffx), then they do so n ther response. They can ndcate ts success wth the approprate status code: HTTP 206 Partal Content [4]. Along wth the status code, they nclude the chunk tself n ther responses. For smplcty, we assume that the clents always request playback from the begnnng of the content and that prefxes are always aalable n the content routers. The content router can ntercept clent requests and deler the prefx drectly to clents. It then contacts the CDN serer to ssue a request for the suffx, and clents can, therefore, recee the remanng part of that content by jonng the suffx streams at the content router. The content router wll calculate the transmsson and recepton schedules so that the tme and channel for transmttng and receng the content are determned usng the schedules [5], [8]. For effcent usage of the bandwdth, t s mportant to know of a deo s popularty. There hae been arous studes related to deo popularty. In [6], deo popularty was reported to follow a Zpf dstrbuton wth skew factor 0.27; that s, 80% of the user s demand s for about 20% of the most popular deos and 20% of the user s demand s for the remanng 80% of the most popular deos. Ths fact helps wth the desgn of the effcent delery schemes, whereby we use a broadcast scheme for popular content and a multcast scheme for less popular content. In ths sense, we assume that content popularty follows the Zpf dstrbuton. Furthermore, we assume that nformaton about the popularty of content s aalable by means of statstcs and expectaton. In addton, preous studes explorng the dstrbuton of multmeda fles n CDNs used Zpf dstrbutons to characterze the popularty of the dfferent contents [5], [6]. Although the popularty of content does not exactly ft the Zpf dstrbuton, many researchers stll adopt the Zpf approach to model popularty n CDNs. Wth the aforementoned assumpton, the costs n Fg. 4 are deduced by usng (9) and (5). We assume that costs assocated wth content routers are manly lnked to the delery, rather than the cachng, of content a fact reflected by the trend n eer-decreasng storage costs. We also consder that there are enough channels n the CDN system so that the probablty of runnng out of Vdeo content F() F(2) F(3) F(4) F(5) F(6) F(7) F(8) F(9) F(0) F() L Serer channel F() F() F() F() F() F() F() F() F() F() F() Serer channel 2 F(2) F(2) F(2) F(2) F(2) F(2) F(2) F(2) F(2) F(2) F(2) Serer channel 3 F(3) F(3) F(3) F(3) F(3) F(3) F(3) F(3) F(3) F(3) F(3) Serer channel 4 F(4) F(5) F(4) F(5) F(4) F(5) F(4) F(5) F(4) F(5) F(4) Serer channel 5 F(6) F(7) F(6) F(7) F(6) F(7) F(6) F(7) F(6) F(7) F(6) Serer channel 6 F(8) F(9) F(0) F() F(8) F(9) F(8) F(9) F(0) F() F(8) Fg. 4. Example of a fast broadcast (FB) scheme when partton functon f(n ) and number of serer channels (K = 6) are gen. ETRI Journal, Volume 36, Number 5, October 204 Jeong Yun Km et al. 82 http://dx.do.org/0.428/etrj.4.03.0090
such channels can be neglected. Some system parameters are dentfed from [7] [9] as follows. We use N to denote the number of content types n the system and S as the total number of content routers. The aalable number of multcast channels n the CDN serer s denoted by N c, and L s the length (n mnutes) of the th content, where N. Each request for content arres at content router s ( s S) accordng to a Posson process wth a rate of, s requests/mn. The aggregate requests for content and the oerall external request rate are gen, respectely, by () S s, s N and. (2) III. Effcent Content Delery Scheme wth In- Network Cachng. Multcast Scheme wth In-Network Cachng In the multcast scheme coupled wth n-network cachng, as shown n Fg. 2, let W be the prefx length for content, whch also corresponds to the patchng wndow for n-network cachng n content routers [5]. Suffxes (of length L ) of content are stored and delered from the CDN serer by means of multcast channels, whle prefxes (of length W ) stored n content serers are delered to clents through uncast streams. When the frst request arres n the content router n steps 3 and 4, a patchng wndow wll be started for tme nteral W. The requests for the same content that arre wthn the wndow wll form a group, and then a sngle multcast from the CDN serer s ntated by the frst request and carred out to all clents n the group. Furthermore, snce the range of the suffx always ncludes that of the prefx, the content router relays the suffx request to the CDN serer n step 6, whereas n response to step 3, t ssues the prefx response to the clent wth an HTTP 204 No Content. Wth an HTTP 200 OK, the CDN serer mmedately begns transmttng the suffx to the content router n step 7, where a copy of the suffx s transmtted to clents wth an HTTP 200 n step 8. For the followng requests that arre later than the frst request, the clents can obtan the mssng ntal porton through a patchng stream wth an HTTP 206 n step 5. At the same tme, they wll obtan the rest of the content by tunng to an ongong multcast stream wth an HTTP 206 n step 8. Once clents start to recee the content from a multcast channel, a patchng stream wll be released after receng the mssng part that the CDN serer cannot transmt to the clent. The patchng stream s, therefore, transent n nature and of a short duraton. For requests for the same content wthn the wndow, the content router repeatedly copes the suffx n proporton to the number of requests and then transmts t to the clents [4]. For the frst request that arres after the end of the patchng wndow, t ntates a new wndow whereby the same operatons should be repeated. Therefore, the aerage nteral between successe multcast streams s gen by W + /. The requred number of multcast channels for the th content s gen by L M,. (3) W / Snce the expected prefx length of the patchng stream s W /2, the total aerage number of channels allocated to the content routers s gen by U W. (4) 2 N M The problem of mnmzng the total aerage number of channels allocated to the content routers s soled by determnng the optmal alue of W, subject to the constrant N M Nc. Snce L s the length of content, we always hae L W 0, and then M 0 from (3). Gen poste constants, the followng optmzaton problem s formulated: ( P) mn, W 2 subject to M N, M 0, N. c The optmzaton problem (P) has a unque optmal soluton that can be obtaned analytcally. It follows from (3) that L W, M. (6) By substtutng (5) for (6), the problem (P) can be rewrtten as L ( P2) mn ( ), 2 M subject to M N, M 0, N. c When the Karush Kuhn Tucker (KKT) condton of (P2) s gen, we can sole the optmal prefx length by settng (P2)/M = 0 and usng the Lagrangan multplers wth respect to the equalty constrant and nequalty constrants. In our system model, we dered the optmal prefx length n (8) that mnmzes the aerage number of channels allocated to the content routers for each content from (P2). The optmal prefx length, whch ndcates the n-network cachng sze n the content routers, s gen by W L L N k k k N c (5) (7). (8) 822 Jeong Yun Km et al. ETRI Journal, Volume 36, Number 5, October 204 http://dx.do.org/0.428/etrj.4.03.0090
From (3), we fnd that there s a trade-off between the prefx length and the number of multcast channels because hang longer prefxes reduces the necessary number of multcast channels of the CDN serer but ncreases the number of uncast channels of the content router. By combnng (4) and (8), when n-network cachng sze W s cached n the content routers, the total aerage number of channels allocated to the content routers for content s gen by N N L k k L k U M. (9) 2 Nc 2 2. Broadcast Scheme wth In-Network Cachng Broadcast schemes, n general, are wasteful when the arral rate s not hgh enough, snce a broadcast channel s scheduled ndependent of any user request and dedcated to a deo content [7], [20] [2]. On the other hand, a broadcast scheme coupled wth n-network cachng, as shown n Fg. 3, not only sgnfcantly reduces the CDN serer and network resource requrements but s also capable of mmedately prodng serce to a large number of clents by takng adantage of nnetwork cachng aalable at the content routers. Before ntatng the requests to the content routers, the CDN serer perodcally broadcasts deo content to the content routers through a number of dedcated broadcast channels, as shown n Fg. 3. When the frst request arres n the content router n steps 3 and 4, t mmedately jons an approprate broadcast channel wthout watng for the begnnng of the next broadcast perod n step 6. Wth an HTTP 206 OK, the content router mmedately begns transmttng a copy of the suffx to the clent (step 7). At the same tme, the content router sends a response ncludng the mssng prefx of the deo content wth an HTTP 206 to the clent (step 5). For the subsequent requests, the same operatons should be repeated as such. Once clents start to consume the content from a broadcast channel, a patchng stream wll be released and the clent keeps playng the remanng part from the broadcast channel. FB s chosen to broadcast the deo content n the system model because of the smplcty of the control system among broadcast schemes. The FB model [7], [2] has been ntroduced to address the scalablty ssue of deo content delery. The scheme makes the serer I/O bandwdth usage ndependent of the number of clents at the expense of a bounded user watng tme. The partton functon f(n ), used to partton the deo content nto some segments, represents the relate length of each segment for content. The FB ddes the deo content nto a geometrcal seres of (, 2, 4,, 2 n ), where n s the number of broadcast channels for content at the CDN serer [7], [20]. We assume that the network bandwdth on the clent sde s only suffcent to support two channels at the same tme. It s the same condton n the case of the multcast scheme. To satsfy ths condton, the partton functon f (n ) of an FB s slghtly modfed by n, 2, 3, f( n) 2 n 4, 5, (0) 2 f( n 2) n 5. An example of an FB scheme s shown n Fg. 4, where partton functon f(n ) and number of serer channels (K = 6) are gen. Channel broadcasts the frst segment F() perodcally, Channels 2 and 3 perodcally broadcast segments F(2) and F(3), respectely. Channels 4 and 5 perodcally broadcast the next two segments; that s, F(4), F(5) and F(6), F(7), respectely. Channel 6 perodcally broadcasts the next four segments; that s, F(8), F(9), F(0), and F(). The length of each segment s F for content. By addng two ntal segments, a clent can jon only one broadcast channel whle receng the patchng stream from the content router. For smplcty of exposton, we defne the summaton of the partton functon h(n ) when the number of the serer channel s K for content. n, 2 n 2, K hn ( ) f( n) 3 3, n n ( n 4)/2 (2 6) n 3, n mod 2 0, ( n 5)/2 (2 8) n 3, n mod 2. () Consder deo content whose length s L. Gen the partton functon f(n ), suppose the number of broadcast channels at the CDN serer K s dedcated to broadcast deo content and let F denote the length of the frst broadcast segment at the content routers. From the defnton of the partton functon, we then hae K L F f( n ) F h( n ). (2) n By settng the frst segment of the suffx broadcast equal n sze to the prefx length, the bandwdth usage on the long-haul path can be substantally reduced [7], [20]. From (2), we can see that there s a trade-off between the number of broadcast channels and the length of the frst segment (that s, n-network cachng sze), snce a smaller number of dedcated CDN serer channels, K, wll result n a larger frst broadcast segment, F. To mnmze the aerage number of channels allocated to content routers, the length of frst segment (that s, n-network cachng sze) should be mnmzed. Ths leads to the followng ETRI Journal, Volume 36, Number 5, October 204 Jeong Yun Km et al. 823 http://dx.do.org/0.428/etrj.4.03.0090
optmzaton problem: ( P3) mn U F, B 2 K Nc K N subject to, 0,. (3) Usng the trade-off between the frst segment length, F, and the number of CDN serer channels, K, (P3) s rewrtten by ( P4) mn UB, L 2 hk ( ) subject to K N, K 0, N. c When the KKT condton of (P4) s gen, we can sole the optmal cachng sze F by settng (P4)/F = 0 and usng the Lagrangan multplers wth respect to the equalty constrant. One of these channels transmts only the frst segment of the deo content. The other channels transmt the remanng segments through ther dedcated broadcast channels. The number of concurrent accesses to a CDN serer s lmted by the number of supportable multcast streams, K. From (P4), the number of dedcated channels of the CDN serer that mnmze the length of the frst segment s then gen by K Nc /2 / 2log[ 2 ( ) ]. N 2 (4) By combnng (2) and (3), when n-network cachng sze F s cached on the content routers, the total aerage number of channels allocated to the content routers for content s gen by (5) and depends on the number of CDN serer channels. U B L K, 2 L K 2, 2 2 L K 3, (5) 2 3 L K 3, mod 2 0, ( K 4)/2 K 2 (2 6) L K 3, mod 2. ( K 5)/2 K 2 (2 8) 3. Adapte Scheme Based on Content Popularty For effcent content delery, t s mportant to know the popularty of the content n queston. We assume that content, ranked accordng to popularty, can be dded nto two groups; the content hang mean arral rates, 2,,, respectely, where N denotes the rank ndex of popularty. Snce a broadcast scheme s scheduled ndependent of any Gen number of serer channels, N c and number of content types, N Determne number of channels and types allocated to broadcast, l and k for all content request do f cost of broadcast, U B < cost of multcast, U M then k = k + end f end for for all content request k do l = l + K end whle Determne number of channels and types allocated to multcast, N c l and N k Content 0 k wth number of channels l belong to Broadcast Content k+ N wth number of channels N c l belong to Multcast Fg. 5. Selecton algorthm for determnng the most sutable delery scheme. user request, the most popular content s lkely to be transmtted through perodc broadcastng. On the other hand, the least popular content s, preferably, transmtted through multcastng because a multcast scheme wll be scheduled only when the content s requested [2]. Therefore, the broadcastng of each deo content demands one or more channels dedcated to t, whle the deo content delered through multcastng usually share a pool of channels of the CDN serer. Owng to the skewed popularty, een among the most popular deo content, a CDN system needs to be desgned for carefully selectng an approprate content delery scheme and ntellgently allocatng resources between the content routers and CDN serer. To account for the skewed popularty, we propose an effcent content delery technque, called a popularty-based adapte content delery scheme, that selects ether a broadcast scheme or a multcast scheme by consderng content s popularty. The proposed adapte content delery scheme broadcasts the most popular content usng the broadcast scheme, whle delerng the least most popular content usng the multcast scheme. Gen the total number of aalable channels (the capacty) of the CDN serer, dstrbutng them for nddual broadcastng and the multcastng pool so as to achee the optmal content delery cost s a nonlnear optmzaton problem. The popularty-based adapte scheme ams to mnmze the aerage total number of uncast channels and the aerage cachng sze of the content routers, usng dynamc programmng, for a group of deo content wth hghly skewed popularty. Dependng on the relate popularty of the content, the adapte content delery scheme selects the most sutable delery scheme for all content, and then t allocates the approprate number of channels to each. By takng adantage of the selecton algorthm for determnng the most sutable delery scheme (see Fg. 5), the proposed adapte scheme classfes the N peces of content 824 Jeong Yun Km et al. ETRI Journal, Volume 36, Number 5, October 204 http://dx.do.org/0.428/etrj.4.03.0090
nto two groups accordng to ther popularty; namely, the most popular content (0 k N ) and the least popular content (N k). The former group s assgned 0 l N c channels for fast broadcastng, and the latter group s assgned the remanng N c l channels for multcastng. Note that one of these groups wll not exst f k = 0, N. Once the specfc alues of k and l are calculated usng the selecton algorthm, the number of broadcast channels and multcast channels are determned by replacng N and N c wth k and l n (9) and (4). By applyng ether a multcast scheme or a broadcast scheme n consderaton of content popularty, the mnmum aerage number of channels of the content routers for the proposed adapte scheme can then be acheed usng the followng dynamc programmng formulaton (P5): k L L j L N j k j ( P5) mn, 0 k N k 2 hk ( ) 2 0 l N Nc l c k subject to K,, 0, 0,. l M k Nc l K M N (6) IV. Performance Analyss In ths secton, we ealuate the performance of the proposed content delery scheme, comparng to a multcast and a broadcast scheme wth n-network cachng. As many researchers [3], [22] hae only showed performance gans oer the core network for the ntroducton of content routers wth nnetwork cachng and dfferent delery schemes, we focus on performance from the perspecte of n-network cachng sze, the number of streamng channels of content routers, and the number of streamng channels of the CDN serer. The performance analyss s based on the followng system parameters: s = 0, N = 200, N c = 800 to,000, L = 90 mn, N = 500 requests/mn, and /( j / j ) requests/mn for =, 2,, N. The relate popularty of the content follows a Zpf dstrbuton wth a skew factor of = 0.27. The aboe system parameters are stll effecte unless noted otherwse. Wthout loss of generalty, let > j for < j N ; that s, content popularty decreases n accordance wth the ndex. Here, the rank ndexes and N denote the most- and least-popular, respectely. The rankng ndex of content popularty N ndcatng the popularty, s used on the x-axs nstead of the arral rate,, snce t can help to clearly understand the dfferent n-network cachng sze on the y-axs. The alues on the x-axs n the followng fgures ndcate the rankng ndex of the content popularty, correspondng to arral rate n Fgs. 6 8. Fgure 6 compares the optmal aerage n-network cachng Aerage n-network cachng sze [W (mn)] 35 30 25 20 5 0 5 800 channels 900 channels,000 channels 0 0 20 40 60 80 00 20 40 60 80 200 Popularty () Fg. 6. Aerage n-network cachng sze nsde content routers a a multcast scheme for dfferent number of CDN serer channels. Aerage n-network cachng sze [W (mn)] 45 40 35 30 25 20 5 0 5 800 channels 900 channels,000 channels 0 0 20 40 60 80 00 20 40 60 80 200 Popularty () Fg. 7. Aerage n-network cachng sze nsde content routers a a broadcast scheme for dfferent number of CDN serer channels. sze of a multcast scheme for dfferent numbers of CDN serer channels (N c = 800, 900, and,000), leadng to a mnmzaton of the number of uncast patchng channels allocated to content routers. The number of CDN serer channels s chosen wthn the range of the aforementoned N c alues to clearly dfferentate the performance of multcast and broadcast schemes, snce the latter always outperforms the former when N c s larger than,00. As the content popularty decreases, a larger cachng sze s gradually needed. The cachng sze changes from 5 (mn) to 33 (mn) for dfferent numbers of CDN serer channels at arral rate = 500 (requests/mn). Wth delerng the cachng porton of the least popular content from content routers, the requred capacty of the CDN serer for the least popular content s reduced. The gan can, therefore, ETRI Journal, Volume 36, Number 5, October 204 Jeong Yun Km et al. 825 http://dx.do.org/0.428/etrj.4.03.0090
Aerage n-network cachng sze [W (mn)] 40 35 30 25 20 5 0 5 800 channels 900 channels,000 channels 0 0 20 40 60 80 00 20 40 60 80 200 Popularty () Fg. 8. Aerage n-network cachng sze nsde content routers a the popularty-based adapte content delery scheme for dfferent number of CDN serer channels. be used to deler the more popular content. On the other hand, the cachng porton of the most popular content decreases as the number of CDN serer channels ncreases. From the aboe obseraton, we dentfy that a trade-off exsts between the capacty of the content router and the CDN serer. Fgure 7 shows the optmal aerage n-network cachng sze of a broadcast scheme for dfferent numbers of CDN serer channels, mnmzng the number of uncast patchng channels allocated to content routers. Smlar to a multcast scheme, the cachng sze ncreased n step-up style. The cachng sze changes from (mn) to 45 (mn) for dfferent numbers of CDN serer channels at arral rate = 500 (requests/mn). Compared to a multcast scheme, the cachng sze s smaller for content of hgh popularty but s larger for content of low popularty. The largest occurrng cachng sze, F = 45 (mn), was equal to half of ts content s playback tme. The storage capacty of content routers s manly occuped by the least popular content. Fgure 8 llustrates the optmal aerage n-network cachng sze of the popularty-based adapte content delery scheme for dfferent numbers of CDN serer channels, mnmzng the number of uncast patchng channels allocated to content routers. We obsere that the proposed adapte scheme requres a total aerage storage of 3,58 (mn), whereas the multcast scheme requres 3,939 (mn) and the broadcast scheme requres 3,265 (mn) for all content when the number of CDN serer channels s,000. The proposed adapte scheme mproes the requred storage capacty of the content routers compared to the multcast and broadcast schemes by about 9% and 3%, respectely. From Fg. 5 and (6), the proposed adapte scheme swtches oer from a broadcast scheme to a multcast scheme when popularty rank ndex s between 54 Total aerage number of uncast patchng channels (U) 4,500 4,000 3,500 3,000 2,500 2,000 Multcast scheme Broadcast scheme Proposed adapate scheme,000 800 850 900 950,000,050,00 Number of CDN serer channels (N c ) Fg. 9. Comparson of the aerage number of uncast patchng channels for dfferent numbers of CDN serer channels. Total aerage n-network cachng sze (mn) 4,500 4,000 3,500 3,000 2,500 2,000 Multcast scheme Broadcast scheme Proposed adapate scheme,000 800 850 900 950,000,050,00 Number of CDN serer channels (N c ) Fg. 0. Aerage n-network cachng sze nsde content routers a a multcast scheme for dfferent number of CDN serer channels. and 200. We can, therefore, achee the optmal n-network cachng sze when applyng the proposed adapte scheme snce the cachng sze of the broadcast scheme suddenly ncreases from popularty ndex rank 54, compared to that of the multcast scheme. The performance of the proposed scheme s compared for all content n terms of the aerage numbers of channels allocated to the content routers, as shown n Fg. 9. The proposed adapte scheme requres an aerage of 2,236 channels at the content routers, whereas the multcast and broadcast schemes requre 3,522 and 2,270 channels, respectely. By applyng the proposed adapte scheme, we can reduce the requred number of channels compared to the other schemes by up to 36%. Fgure 0 llustrates a comparson of the aerage total nnetwork cachng sze allocated to the content routers for 826 Jeong Yun Km et al. ETRI Journal, Volume 36, Number 5, October 204 http://dx.do.org/0.428/etrj.4.03.0090
Percentage of broadcast channels (%) 00 95 90 85 80 75 70 Skew factor = 0.27 Skew factor = 0.22 Skew factor = 0.7 65 800 850 900 950,000,050,00 Number of CDN serer channels (N c ) Fg.. Fracton of content delered a broadcast channels for dfferent skew factor n adapte content delery scheme. dfferent numbers of multcast channels of the CDN serer. The aerage total cachng sze of the adapte scheme s close to that of the multcast scheme when the number of channels of the CDN serer s the smallest; that s, at N c = 800. Otherwse, when t gradually ncreases, we obsere that the aerage total cachng sze of the proposed adapte scheme s almost close to that of the broadcast scheme. The fracton of content delered a broadcast channels for dfferent skew factors n the adapte content delery scheme s shown n Fg.. The fractons are dstrbuted ery smlar to each other, regardless of the dfferent skew factors, when the number of channels of the CDN serer s between 800 and 900. On the other hand, when the number s aboe 950, the fractons are dstrbuted wth more and more dersty as the skew factor ncreases. In partcular, the fractons approach 97% when skew factor s 0.27. The results of the performance analyss n ths secton show that the adapte scheme consderably outperforms other schemes by consderng the content popularty, snce the most popular content s delered through broadcast channels and the least popular through multcast channels. V. Concluson Ths paper proposed the popularty-based adapte content delery scheme n a hybrd CDN system that takes adantage of the tradtonal CDN serer n the oerlay and noel content routers n the underlay, whle adoptng n-network cachng n the content routers. By employng the proposed scheme, content routers can adaptely select the most sutable delery scheme and allocate the approprate number of channels to effcently mnmze both ther streamng and storage capactes for all content, dependng on the relate popularty. We showed that the proposed scheme prodes a notable performance gan aganst both the multcast and broadcast schemes coupled wth n-network cachng n terms of the optmal n-network cachng sze and number of uncast channels n a content router. References [] D.D. Vleeschauwer and D.C. Robnson, Optmum Cachng Strateges for a Telco CDN, Bell Labs Tech. J., ol. 6, no. 2, Sept. 20, pp. 5 32. [2] K. Cho et al., How Can an ISP Merge wth a CDN?, IEEE Commun. Mag., ol. 49, no. 0, Oct. 20, pp. 56 62. [3] G. Haßlnger and F. Hartleb, Content Delery and Cachng from a Network Proder s Perspecte, Comput. Netw., ol. 55, no. 8, Dec. 20, pp. 399 4006. [4] W.K.S. Tang et al., Optmal Vdeo Placement Scheme for Batchng VOD Serces, IEEE Trans. Broadcast., ol. 50, no., Mar. 2004, pp. 6 25. [5] B. Wang et al., Optmal Proxy Cache Allocaton for Effcent Streamng Meda Dstrbuton, IEEE Trans. Multmeda, ol. 6, no. 2, Apr. 2004, pp. 366 374. [6] L. Gao and D. Towsley, Threshold-Based Multcast for Contnuous Meda Delery, IEEE Trans. Multmeda, ol. 3, no. 4, Dec. 200, pp. 405 44. [7] L. Gao, J. Kurose, and D. Towsley, Effcent Schemes for Broadcastng Popular Vdeos, Multmeda Syst., ol. 8, no. 4, July 2002, pp. 284 294. [8] S.H. Gary Chan, Operaton and Cost Optmzaton of a Dstrbuted Serers Archtecture for on-demand Vdeo Serces, IEEE Commun. Lett., ol. 5, no. 9, Sept. 200, pp. 384 386. [9] Van Jacobson et al., Networkng Named Content, Proc. CoNEXT, Tokyo, Japan, Dec. 20, pp. 2. [0] D. Eager, M. Vemon, and J. Zahorjan, Mnmzng Bandwdth Requrements for on-demand Data Delery, IEEE Trans. Knowl. Data Eng., ol. 3, no. 5, Sept. Oct. 200, pp. 742 757. [] J.Y. Km, G.M. Lee, and J.K. Cho, Effcent Multcast Schemes Usng In-Network Cachng for Optmal Content Delery, IEEE Commun. Lett., ol. 7, no. 5, May 203, pp. 048 052. [2] A. Barbr et al., Known Content Network (CN) Request-Routng Mechansms, RFC3568, July 2003. [3] M. Masa and E. Parracn, Impact of Request Routng Algorthms on the Delery Performance of Content Delery Networks, IEEE Int. Performance, Comput. Commun. Conf., Apr. 9, 2003, pp. 5 2. [4] S.A. Thomas, HTTP Essentals, Hoboken, NJ: John Wley & Sons, 200. [5] I. Psaras et al., Modellng and Ealuaton of CCN-Cachng Trees, Proc. IFIP Netw., Valenca, Span, 20, pp. 78 9. ETRI Journal, Volume 36, Number 5, October 204 Jeong Yun Km et al. 827 http://dx.do.org/0.428/etrj.4.03.0090
[6] J. Cho, A.S. Reaz, and B. Mukherjee, A Surey of User Behaor n VoD Serce and Bandwdth-Sang Multcast Streamng Schemes, IEEE Commun. Sureys Tutorals, ol. 4, no., 202, pp. 56 69. [7] G. Xue, Serer Cost Mnmzaton n a Dstrbuted Serers Archtecture for on-demand Vdeo Serces, IEEE Commun. Lett., ol. 7, no. 9, Feb. 2003, pp. 52 54. [8] D. Guan and G. Xong, Optmal Prefx Cache Allocaton among Multple Cooperate Local Proxes, Int. Conf. Wreless Commun. Netw. Moble Comput., Bejng, Chna, Sept. 24 26, 2009, pp. 4. [9] L. Dong et al., Performance Ealuaton of Content Based Routng wth In-Network Cachng, Wreless Opt. Commun. Conf., Newark, NJ, USA, Apr. 5 6, 20, pp. 6. [20] L. Gao, Z. L. Zhang, and D. Towsley, Proxy-Asssted Technques for Delerng Contnuous Multmeda Streams, IEEE/ACM Trans. Netw., ol., no. 6, Dec. 2003, pp. 884 894. [2] S.A. Azad and M. Murshed, An Effcent Transmsson Scheme for Mnmzng User Watng Tme n Vdeo-on-Demand Systems, IEEE Commun. Lett., ol., no. 3, Mar. 2007, pp. 285 287. [22] Y. Km and I. Yeom, Performance Analyss of In-Network Cachng for Content-Centrc Networkng, Comput. Netw., ol. 57, no. 3, Sept. 203, pp. 2465 2482. Jeong Yun Km receed hs BS and MS degrees n electronc engneerng from Inha Unersty, Incheon, Rep. of Korea, n 990 and 992, respectely and receed hs PhD degree n nformaton and communcaton engneerng from the Korea Adanced Insttute of Scence and Technology, Daejeon, Rep. of Korea, n 204. Snce 992, he has been wth the Electroncs and Telecommuncatons Research Insttute, Daejeon, Rep. of Korea as a specal fellow. Hs man research nterests are Future Internet, streamng serces, and energy sang technologes ncludng smart grds. He has actely partcpated n standardzaton meetngs ncludng ITU-T SG 3 (Future Networks & Cloud) as an edtor and IETF. He has contrbuted more than 200 proposals for standards and publshed more than 50 papers n academc journals and conferences. He s a member of the IEEE. Gyu Myoung Lee receed hs BS degree n electronc and electrcal engneerng from Hong Ik Unersty, Seoul, Rep. of Korea, n 999 and hs MS and PhD degrees from the Korea Adanced Insttute of Scence and Technology (KAIST), Daejeon, Rep. of Korea, n 2000 and 2007. In 2007, he worked as a guest researcher at the Natonal Insttute of Standards and Technology, Gathersburg, MD, USA. Later that year, he was nted to work on the research staff at the Electroncs and Telecommuncatons Research Insttute, Daejeon, Rep. of Korea. In 2008, he was wth the Insttut Mnes-Telecom, Telecom SudPars, Ery, France, as an adjunct assocate professor. Then n 202, he contnued hs work as an adjunct professor at KAIST, Daejeon, Rep. of Korea. Recently he has been employed as a Senor Lecturer at the Lerpool John Moores Unersty, Lerpool, UK. Hs research nterests nclude Internet of thngs, future networks, multmeda serces, and energy sang technologes ncludng smart grds. He has actely partcpated n standardzaton meetngs, ncludng ITU-T SG 3 (Future Networks & Cloud) as a rapporteur, onem2m, and IETF. He has contrbuted more than 300 proposals for standards and publshed more than 00 papers n academc journals and conferences. He s a senor member of IEEE. Jun Kyun Cho receed hs BS degree n electroncs from Seoul Natonal Unersty, Seoul, Rep. of Korea, n 982, and hs MS and PhD degrees from the Korea Adanced Insttute of Scence and Technology (KAIST), Daejeon, Rep. of Korea, n 985 and 988, respectely. He worked for ETRI from 986 to 997 and s currently workng as a professor at KAIST. 828 Jeong Yun Km et al. ETRI Journal, Volume 36, Number 5, October 204 http://dx.do.org/0.428/etrj.4.03.0090