JP2010057107A - Server disposition method, server disposition method in carrier type cdn, server disposition system, carrier type cdn system, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To minimize a total traffic volume flowing inside a carrier type CDN (Contents Delivery Network). <P>SOLUTION: A server base disposition apparatus 101 selects one of nodes on a network, disposes an original server system for storing all contents (S901), selects a plurality of nodes from among the nodes, and disposes them as a content server 103 (cache server system) for caching and distributing contents stored in the original server system (S902). In such a case, the original server system and the cache server system are selected so as to minimize an average hop count of a distribution flow by input of, as input information, an NW topology and a generation frequency of distribution requests from each client computer to the contents. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to content distribution server placement technology on the Internet, and more particularly to server placement technology suitable for efficiently delivering large-capacity rich content in a carrier-type CDN (Contents Delivery Network). .

  Conventionally, on the Internet, downloading of small-sized contents with a small capacity such as texts such as Web and still images has been the main usage form. However, in recent years, streaming distribution of very large moving image content of several hundred megabytes to several gigabytes or more such as movies and live broadcasts provided by various content providers is rapidly spreading.

  Therefore, the Internet is required to distribute these large-capacity rich contents economically and comfortably without stressing the user and without burdening the network (NW).

  Conventionally, CDN has been widely used as a technique for distributing small-sized contents to users. There are many CDN providers that provide this CDN service to content providers, including Akamai Technologies (trademark) and Lifetime Networks (trademark).

  CDN is a technology designed in an era when the transmission capacity of the NW infrastructure was limited, and aims to improve the response time of users, placing as many servers as possible on various NWs and being close to the users Use the server for distribution.

  For example, in the case of Akama Technologies (trademark), more than 20,000 servers are arranged on about 1,000 NWs in 70 countries around the world.

  Thus, since the number of servers is enormous, the capacity of each server is limited. Conventionally, contents distributed on the Internet are mainly small-sized contents such as text and still images, and such distribution forms are also effective.

  The above prior art is described in Non-Patent Document 1 and the like.

  In recent years, the use of video content streaming distribution services has increased rapidly. For example, in the case of a system for distributing video content (UGC: User-Generated Content) created by the user himself such as YouTube (trademark), the capacity (size) of each content is small, and thus a conventional CDN can be sufficiently handled.

  However, in the future, it is expected that the distribution of attractive content owned by leading content providers such as TV broadcasting stations and movie production companies will spread on the Internet. These contents have a large capacity, and it is difficult to provide such a large-sized content (rich content) with a conventional CDN because the number of contents that can be stored in each server is greatly limited.

  In order to efficiently deliver such rich contents, Light Networks Networks (trademark) has established a large-scale data center (computer system) in a limited location and connected them with its own large-capacity NW infrastructure. We are starting to develop a new form of CDN business.

  This new form of CDN is called “second generation CDN”, and each data center aims to integrate a large number of servers with clustering technology and prepare a large amount of rich content in all data centers. ing.

  In recent years, P2P (Peer to Peer) type content distribution such as BitTorrent (registered trademark) receiving distribution from other users who possess the target content has become widespread.

  For example, in BitTorrent (registered trademark), content data is divided into a plurality of blocks, and a user who wants to acquire content downloads the block and at the same time, another user (peer) who wants to distribute the same content. Upload in parallel.

  In this technology, as the number of users who want to acquire content increases, the number of peers holding the corresponding content increases. Therefore, it becomes a very effective distribution form when distributing extremely popular content to a large number of users. .

  Normally, when there are a large number of servers having the same content, it is desirable to receive distribution from a server that exists at a position with a small number of via hops from the viewpoint of NW consumption resource amount and user response time.

  Therefore, in many CDNs, distribution is performed from a server with a short hop distance in response to a user's distribution request.

  However, in many P2P type distribution systems such as BitTorrent (registered trademark), since a peer as a distribution source is randomly selected, even when a peer exists nearby, a peer that exists far away is selected and distributed. In some cases, traffic flows wastefully in the NW.

  In recent years, the increase in UGC distribution using P2P type distribution has been remarkable, and resource waste in the NW has become a problem. Therefore, P4P (Pay for Performance) is being studied in which ISP (Internet Services Provider) and P2P distributors exchange information with each other to perform peer selection while suppressing the amount of traffic flowing in the NW.

  This P4P enables a P2P type distribution service that suppresses the influence on the NW, but the peer that plays the role of the distribution server is the user, and the behavior is selfish, so that It is necessary to give motivation for uploading.

  In addition, since the positions of all servers cannot be designed and managed in a centralized manner, the servers cannot be arranged at the optimum positions when viewed from the entire NW, and the impact on the NW compared to conventional P2P. However, it cannot be said that an optimal distribution form is realized.

  Further, since there is a possibility that the server that is being distributed may be disconnected, it is not suitable for streaming distribution of moving image content for which continuous and stable distribution is desired.

  In addition to P4P, an approach to reduce the load on the CDN server by using P2P in a complementary manner in combination with P2P and CDN has been studied, but the main purpose is to reduce the server cost of CDN. It is not aimed at the realization of a distribution system that minimizes the impact on the network.

  In general, the position of the user that is the receiving side of the content distribution flow generated by the distribution request of each user cannot be changed, but the position of the server that is the transmitting side is flexible and can be freely designed and controlled. It becomes a big factor that determines the exchange of traffic flowing in.

  Therefore, when delivering rich content that has a large capacity and has a large influence on the NW of the delivery flow, it is possible to centrally manage the location of the server and the content, and deliver from the optimum server for each delivery request. is important.

  As one method for that purpose, a carrier (telecommunications carrier) or ISP establishes a large capacity server base (data center) at a plurality of locations in its own NW, and constructs a CDN that distributes rich content by itself. Is promising.

  Hereinafter, such a form is referred to as a “carrier type CDN”. In this carrier type CDN, the carrier and ISP can centrally manage all server bases in addition to information about the NW infrastructure such as the topology and link capacity of the own NW.

  Furthermore, in this carrier-type CDN, unlike a CDN operator that arranges a myriad of small-capacity servers in the past, a content center is constructed by accommodating a large-capacity server group in a limited number of locations. This makes it easy to centrally manage the location where the copy exists, and also allows a large number of rich contents to be prepared at a large number of server locations, thereby improving the distribution opportunity of the rich contents from the optimal position.

  In fact, AT & T (TM) has recently announced that it is developing a CDN business. In addition, Lifelight Networks (trademark) has built an NW infrastructure by itself, and the second generation CDN operator can be regarded as a carrier, and can be considered to be a form close to a carrier type CDN.

  As described above, the carrier-type CDN is promising as a distribution system that performs the distribution of rich content under an optimal distribution design that suppresses the influence on the NW.

  However, no technology has been proposed for optimally arranging data centers in such a carrier type CDN distribution system.

M. Karlsson and M. Mahalingam, “Do We Need Replication Placement Algorithms in Content Delivery Networks?” WCW 02.

  The problem to be solved is that the conventional technology cannot optimally arrange the data center in the distribution system using the carrier type CDN.

  The object of the present invention is to solve these problems of the prior art and optimize the original server base and the cache server base to minimize the average hop length of the distribution flow and the average load of each link in the carrier type CDN. It is possible to arrange.

In order to achieve the above object, in the present invention, for example, in a carrier type CDN, the original server base and the cache server base are optimally arranged so as to minimize the average hop length of the distribution flow and to minimize the average load of each link. It is characterized by. In other words, the carrier builds a large-capacity data center at a limited base in its network (NW), and in the carrier-type CDN that operates the CDN, one original server base that always accommodates all content is arranged. Further, based on a distribution request from the user, for example, cache server bases for updating the content by LFU (Least Frequency used) are arranged in a plurality of data centers. For example, when an NW topology and a frequency of distribution requests for each content are given from each node, one original server base and an arbitrary number of cache server bases are set so that the average hop length of the distribution flow is minimized. Optimal placement. In particular, for each node n, the centrality H _o (n) of the node n is defined by the reciprocal of the sum total of the number of hops on the shortest path from the node n to each node weighted by the request occurrence ratio. Place the original server on the largest node. Further, the cache server capacity K of the entire NW is given, and the capacity B of each cache server is given as a fixed value (B = floor (K / C)) with respect to the number C of the arranged cache server bases. Including the optimal placement of cache server bases. More specifically, h _{n, min} and h _{n, o} are respectively the hop distance from the node n to the shortest hop distance among all S server bases including the original server base, and the original server the hop distance to the node n, further distribution request Q n _(m) is from the node n is the cumulative distribution function of a probability q _{n (m)} is intended for content m, and, of r _n from node n When the request generation ratio is set, the average hop length of the distribution flow is estimated by the following equation (1).

また、ｈ_{ｉ，ｍｉｎ，Ｎｊ}を、Ｎ_ｊにキャッシュサーバ拠点が設置された状態における、ノードｉから最近接サーバ拠点までのホップ長とするとき、既にｊ個のキャッシュサーバ拠点がノード集合Ｎ_ｊに設置された状態で（Ｎ_０は全ノード集合Ｎからオリジナルサーバ拠点Ｏを除いたＮ＼Ｏ；「＼」は集合の引算）、さらに、全ノード集合Ｎから集合Ｎ_ｊを除いた集合におけるノードｎ（ｎ∈Ｎ＼Ｎ_ｊ）にキャッシュサーバ拠点を設置することで得られる配信フローの平均ホップ長の推定値ｈ＾の削減効果Ｅ_ｎ、Ｎｊを、下記の数２により算出し、Ｅ_ｎ、Ｎｊが最大のノードから順にキャッシュサーバ拠点を配置していく。

Further, when h _{i, min, Nj} is the hop length from the node i to the nearest server base in a state where the cache server base is installed at N _j , _j cache server bases are already in the node set N _j (N ₀ is N \ O excluding the original server base O from the entire node set N; “\” is a subtraction of the set), and further, the set excluding the set N _j from the total node set N The reduction effect E _{n, Nj} of the estimated value h ^ of the average hop length of the distribution flow obtained by installing the cache server base at the node n (n∈N \ N _j ) in Cache server bases are arranged in order from the node with the largest _{En and Nj} .

  According to the present invention, in the carrier type CDN, the original server base and the cache server base can be optimally arranged so as to minimize the average hop length of the distribution flow and to minimize the average load of each link. It is possible to minimize the total traffic flowing in the network.

  The best mode for carrying out the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a network configuration example of a carrier type CDN provided with a server arrangement system according to the present invention, and FIG. 2 shows an example of a distribution situation of average node order and maximum node order of 36 networks. FIG. 3 is an explanatory diagram showing an example of topology and node population of Caboe & Wireless, FIG. 4 is an explanatory diagram showing an example of simulation results of the present invention for Caboe & Wireless, and FIG. 5 is an example of design results according to the present invention for Caboe & Wireless. FIG. 6 is an explanatory diagram illustrating a performance comparison example in 18 networks, FIG. 7 is an explanatory diagram illustrating a performance comparison example in 36 networks, and FIG. 8 is a normalization in 36 networks. FIG. 9 is an explanatory diagram showing the number of cache servers, and FIG. It is a flowchart showing a server arrangement procedure example of the present invention by the device.

  In FIG. 1, reference numeral 101 denotes a server location arrangement device, 102 denotes a central control station, 103 denotes a content server, 104 denotes a server selection device, and 105 denotes a user terminal, each of which includes a CPU (Central Processing Unit), a main memory, and a display device And a computer configuration including an input device, an external storage device, etc., and after the program or data recorded on the storage medium such as a CD-ROM is installed in the external storage device via the optical disk drive device, etc., this external storage Programmed computer processing is performed by reading from the device into the main memory and processing by the CPU.

  As programmed computer processing, the server base placement apparatus 101 optimally places server bases (nodes on which server computers for delivering content are placed), and the central control station 102 is informed of distribution requests generated by the user terminal 105. The server selection device 104 requests the server selection device 104 to select the server, and the server selection device 104 that has received the request selects one server with the shortest communication distance from the user terminal 105 and requests the selected content server 103 to distribute In response to the distribution request, the content server 103 distributes the content to the user terminal 105.

  At that time, other content servers existing on the distribution route cache and hold the content based on LFU (Least Frequently Used), and notify the central control station 102 and the server selection device 104 of the information of the held content. To do.

  The carrier-type CDN network shown in FIG. 1 is provided with a plurality of content servers (server computer systems) that distribute content to user terminals 105 (client computers) via the network. A plurality of server computer systems are arranged according to the processing procedure shown in FIG.

  In other words, the server site placement apparatus 101 performs one programmed node process from each node on the network (specifically, a network exchange point in which a device having a router function is installed). A first procedure for selecting and arranging an original server system for storing all contents in the selected node (step S901), selecting a plurality of nodes from each node on the network, and selecting the original server for each selected node A second procedure (step S902) for arranging a cache server system (content server 103) that caches and distributes content stored in the system is executed.

  At this time, in the first and second procedures, the server site placement apparatus 101 inputs the network topology information and the frequency of occurrence of distribution requests for each content from all nodes, and the average hop length of the distribution flow is minimized. The nodes on which the original server system and the cache server system are arranged are selected so that they can be realized.

In particular, in the first procedure, the server site placement apparatus 101 determines, for each node n on the network, the number of hops on the shortest path from the node n to each other node by the ratio of occurrence frequency of distribution requests. The centrality H _o (n) of the node n is defined by the reciprocal of the weighted sum, and the node having the maximum centrality H _o (n) is selected to place the original server.

  Further, in the second procedure, the server site placement apparatus 101 sets the capacity B of each cache server system to a constant value, and uses the capacity K of the cache server system of the entire network and the number C of each cache server system. By using the expression “B = floor (K / C)”, the node where each cache server system is arranged including the number C of each cache server system is selected.

In addition, in the first and second procedures, the server base placement apparatus 101 determines the hop distance h to the node having the shortest hop distance from the node n among the S nodes that place the original server system and the cache server system. _{n, min} , the hop distance h _{n, o} from the node where the original server system is located to the node n, and the cumulative distribution function Q of the probability q _n (m) that the delivery request from the node n is for the content m and _{n (m),} by using the distribution request generation ratio r _n from the node n, calculated from equation (3) below, the estimate of the average hop length distribution flow a (h ^).

In addition, in the second procedure, the server site placement apparatus 101 sets a node N _j where each of the j cache server systems is placed from the node i (where j = 0, the node where the original server system is placed). The hop length h _{i, min, Nj} to the nearest node in the excluded set N ₀ ) _, and the hop length h _{n, i} from the node i to a node n outside the set N _j (nεN \ N _j ) Is obtained by arranging a new cache server system at node n in a state where _j cache server systems are already installed at each node included in the set N _j from the following equation (4). reduction E _n of the average hop length distribution flow (Suichi value) "h _^", calculates _Nj, place the cache server system calculated savings E _n, from _Nj is the largest node in order It will be selected as that node.

  Next, the details of the carrier type CDN and the server arrangement technique in the carrier type CDN according to the present invention will be described.

  First, “3.5 Carrier CDN” will be described.

  In the carrier type CDN, a single carrier or ISP having a large backbone NW such as Tier 1 has a large data center in a limited number of bases (nodes) on the NW operated by the single carrier or ISP. Build up. The number of nodes in the entire NW is N, the number of nodes (server bases) provided with a large-scale data center is S, and the number of rich contents is M.

  In addition, a central control station exists, manages the contents existing at each server base, and selects a server used for distribution in response to each distribution request.

  In this example, it is assumed that the CDN service is closed to a single carrier or ISP, and the purpose is to minimize the total amount of traffic flowing in the NW. However, a carrier type CDN service in which a plurality of carriers and ISPs cooperate with each other. The form which expands can also be considered.

  Each data center accommodates a large number of server devices using clustering technology, and has a very large storage capacity and a simultaneous processing capability for a large number of distribution requests.

  Next, “3.5.1 Content development method” will be described.

  How to distribute the copy of each rich content to S server bases becomes a problem, but in the conventional CDN, the replication method and the caching method have been studied.

  In the replication method, a server set that distributes each content is derived as an optimization problem (RPA: Replication Placement Algorithm) such as minimizing the average number of hops in the distribution flow, and a copy of each content is developed on the corresponding server in advance.

  Various contents development methods are being studied. A number of methods have been studied for the RPA problem, but it is necessary to estimate the frequency of distribution requests generated for each content in each node in advance.

  Since the RPA problem is NP-hard, it is necessary to use an approximate solution, and various methods have been studied. However, the Greedy method in which content is simply arranged at a node having the best degree of improvement of the objective function is the best.

  On the other hand, in the Caching method, content is not distributed to servers in advance, and each server is caused to function as a cache server. That is, the server closest to each distribution request operates as a proxy server, and when there is no corresponding content, it is downloaded from the original server, distributed to the user, and simultaneously caches the corresponding content.

  If the server capacity is exceeded, the cache replacement algorithm such as LFU (Least Frequently Used) or LRF (Least Recently Used) replaces the content with the least access frequency or the longest elapsed time since the last request. To do.

  If content information held by each cache server is exchanged between the servers and shared, it is possible to perform distribution from a neighboring server that holds the corresponding content instead of the original server when a cache miss occurs. Therefore, for the purpose of improving response time and reducing consumption network resources, a method of coordinating between caches has been studied.

  In this caching method, content with a high request frequency is easily cached in each server. Therefore, in a realistic environment in which the access frequency dynamically changes, an ideal caching method that shares information between servers shows better characteristics than an optimal replication method.

  However, when server information is not shared, in an environment where the server capacity is small and the cache miss rate is high, the frequency of distribution from the original server having a high distribution cost is high, and the performance is degraded compared to the replication method.

  Many of the actual CDNs use a caching method that does not share server information, although its performance deteriorates because it is easy to operate.

  As a method for deploying content to servers, it is desirable to use an ideal caching method in which information is shared between servers in order to reduce the traffic volume and response time within the NW.

  By the way, in the carrier type CDN, the server bases are limited to a small number of locations and the existence of a centralized control station that manages the whole is assumed. Therefore, it is possible to centrally manage content information held by each server base. .

  Therefore, in this example, it is assumed that an ideal caching method is used as a content development method in the carrier type CDN.

  However, in order to guarantee delivery of content, at least one copy of each content needs to exist in S server sites. Therefore, in this example, it is assumed that one of the S server bases is the original server base, and that all the M contents are always held in the original server. The other C (= S-1) server bases function as cache servers.

  In this case, in the initial state, the content exists only in the original server, but the content is cached in the cache server in accordance with the user's distribution request, so that the content spreads to all the server bases.

  Even after the service operation starts, it is expected that content providers will provide new rich content in a timely manner, but in this case as well, by uploading the content only to the original server, after that, according to the user's distribution request, The content automatically spreads to other server bases of the NW.

  Next, “3.5.2 Distribution processing sequence” will be described.

  The distribution request for the rich content m generated from the user terminal (105) accommodated in an arbitrary node n is notified to the central control station (102). The central control station (102) keeps track of all the server bases holding the content m, and distributes one of the server bases with the shortest hop distance to the node n, including the original server. Choose as.

  When there are a plurality of objects having the shortest distance, one is selected at random. If no server site caches the content m, the distribution is performed from the original server.

  Then, the central control station (102) instructs the selected distribution server (content server 103) to distribute the content m to the node n. The content distribution route is assumed to be the shortest hop route of Dijkstra. If there are a plurality of shortest hop routes, one is selected at random.

At each server site, the number of times X _j delivered in the past W time is stored in the storage device in association with each content j (identification information thereof), and the cache server is updated by LFU.

If the server m existing on the route from the selected server site to the node n does not hold the content m and the server has insufficient free space, the number of distributions X _m for the content _m is Then, the minimum value (X _min ) of the number of distributions X _j in the possessed rich content is compared, and if X _m > X _min , the content with the smallest X _j is deleted and the content m is stored instead.

  In this example, for simplicity of explanation, the size of all M contents is the same, but can be easily expanded even when the size differs for each content. However, the contents being distributed to the user are excluded from the deletion target. If the server has free space, the content m is simply cached.

  In this way, when a change occurs in the content held, the server base notifies the centralized control station (102) of the update information of the held content.

  Thus, by updating the content to be cached by the LFU, it is possible to always cache popular content while autonomously updating the owned content at each cache server base.

  In addition, when the popularity ranking of content does not change much, it is desirable to set the past time W to be large and to estimate the popularity of each content conservatively. Conversely, when the popularity ranking changes frequently, It may be desirable to set the past time W small to reflect a shorter-term popularity trend.

  Next, “3.6 Server Base Location Design Method” will be described.

  In the conventional CDN, since an infinite number of servers are arranged on the NW, there is almost no examination on the optimal arrangement position of the servers. For example, in the replication method, it is assumed that servers exist in all nodes, We are trying to optimize the layout of content. In addition, there are few studies on the optimal server placement method in the caching method.

  For this reason, assuming a cache replacement algorithm such as LFU, there is no research that deals with the optimal server placement problem for an arbitrary topology when only some contents can be cached.

  On the other hand, in the carrier type CDN of this example, since a large-capacity server group is installed in a limited base, the arrangement of the server base has a great influence on the performance. Therefore, before starting the performance evaluation of the carrier-type CDN, the optimum arrangement method of the server bases will be examined.

  First, “3.6.1 Greedy method design algorithm” will be described.

  In general, since the server allocation problem is an integer programming problem, when an arbitrary topology is targeted, it becomes an NP difficult problem, and it is necessary to use an approximate solution except for a very small NW. Since the Greedy method is the best for many optimal placement problems, a design algorithm based on the Greedy method is also examined in this example.

  When designing optimally from the standpoint of a carrier or ISP, the design goal is to minimize (i) the total traffic flowing in the NW (the sum of the traffic flowing through each link), (ii) the link with the highest utilization rate It can be considered to minimize the usage rate of.

  In this example, the optimal design is considered with the above-mentioned (i) as the target of the design target. (I) is equivalent to minimizing the average hop length of the distribution flow, and it is expected to improve the response time of the user while minimizing the consumed NW resource amount.

The ratio of distribution request number R generated in the whole NW delivery request number _{R n} generated from the node n to be sufficiently long in T and _{r n (r n = R n} / R). Also, let q _n (m) be the probability that the distribution request from the user terminal accommodated in the node n is for the content m. It is assumed that the ratio r _n and the probability q _n (m) are known for all n (nodes) and m (contents).

  Of the S server bases, one is the original server base and the other C (C = S-1) is the cache server base, and the arrangement problem between the original server base and the cache server base is considered separately.

  First, “original server location method” will be described.

  In order to minimize the average hop length of the distribution flow, it is desirable to place the server base at a node located at the center of the NW having a short hop length on the shortest hop route to another node.

Therefore, considering that all the M contents are held in the original server base, the centrality H _o (n) of the node n with respect to the arrangement of the original server base is defined by the following equation (5).

In Equation 5, “h _{n, i} ” is the number of hops on the shortest hop route from the node n to the node i, and h _{n, n} = 0. Then, the original server base is installed at the node having the maximum centrality H _o (n).

  Next, “cache server location arrangement method” will be described.

After determining the original server base, the problem of placing the C cache server bases among the N-1 candidates is a combinatorial optimization problem. To obtain a strict optimal solution, O (2 ^N ) Dimensional complexity is required. Therefore, the calculation time increases exponentially as N increases, and it is difficult to obtain an exact solution for a large-scale NW. Therefore, an approximate optimum design method based on the Greedy method is examined.

  When the size of each content is L (bytes) and the average hop length of the distribution flow is h, the total traffic amount V (bytes) obtained by adding the traffic amount flowing through each link within the period T to all the links is as follows. Equation 6 is obtained.

However, in the formula of h, h _n is the average hop length of the distribution flow for the node n. Even if the position of the server base is changed, R (the number of distribution requests generated in the entire NW) and L (the size of each content) are unchanged, so the average hop length h of the distribution flow obtained by dividing V by R and L Minimize design goal.

  Since the Caching method is assumed as a method for deploying content to server bases, the set of server bases holding each content m varies stochastically. For this reason, the average hop length h when C cache server locations are determined cannot be calculated strictly.

  However, when the cache content is updated autonomously at each cache server site using LFU, since the popular content is cached on average, the estimated value “h ^” of the average hop length h of the distribution flow is set. It is defined by the following equation (7).

However, in the above formula 7, “B” is the server capacity of the cache server base and can hold B contents at the maximum. “H _{n, min} ” and “h _{n, o} ” are the hop distance from the node n to the shortest hop distance among all S server bases including the original server base, The hop distance from the original server to the node n.

Furthermore, in the above formula 7, “Q _n (m)” is a cumulative distribution function of “q _n (m)”, and “Q _n (m) = Σ _{i = 1} ^m q _n (i)”. It is. That is, the C cache server bases are considered to have popular top B content, and the top B content is the remaining “MB” content from the nearest server base. Is delivered to each user from the original server.

  By the way, how to design the number C of cache server bases becomes a problem. If the server capacity B is constant, the estimated value “h ^” of the average hop length h is increased as the number C of server bases increases. And the traffic volume in the NW is reduced. However, on the other hand, the cost required to construct a server base increases.

  Therefore, the server capacity of the entire cache server base is given as a constant value K (unit is the number of contents), and the base number C is included so that the estimated value “h ^” of the average hop length h is minimized under this restriction. Design optimally.

At this time, the server capacity B is a maximum integer value not exceeding “K ÷ C” (a constant value of the server capacity ÷ the number of server bases). Since the server capacity B needs to satisfy “B ≧ 1”, the allowable upper limit C _max of the number of server bases C is “C _max = max (N−1, K)”.

Further, increasing the server capacity B beyond the total number of contents M does not contribute to the improvement of the estimated value “h ^” of the average hop length h, so the lower limit C _min of the number of server sites C is set to “K ÷ M”. Set to the smallest integer value not less than.

  The server installation method using the Greedy method is a method in which servers are sequentially installed in a place where the objective function is most improved, and the once installed servers are not removed.

In a state where _j cache server bases are already installed in the node set N _j (N ₀ is N \ O excluding the original server base O from the total node set N), and further a node n (nεN \ N _j ) If the reduction effect of the estimated value “h ^” of the average hop length obtained by installing the cache server base in “E _{n, Nj} ” is _expressed by the following equation (8).

In the equation of Equation 8, "h _{i, min, Nj"} is in a state in which the cache server site is installed in the node set N _j, is the hop length to the nearest server based from node i.

However, although it is necessary to multiply by “Q _i (B)” in the formula (8), it will be multiplied regardless of the selection location of the node n.

  The following (I) to (IV) summarize the cache server location optimal placement algorithm based on the Greedy method for minimizing the estimated value “h ^” of the average hop length of the distribution flow.

(I) Initial setting to k = C _min , N ₀ = N \ O, j = 0.
(II) The reduction effect E _{n, Nj} is calculated by the formula 8 for all nodes n of n∈N \ N _j .
(III) Reduction server E _{n, Nj} has the largest cache server base at node n * and is updated to “j = j + 1” and “N _j = N _j + n *”.
(IV) If j <k, proceed to (II). If j = k, the estimated average hop length “h _k ^” in the optimal cache server base arrangement of C = k Calculate, k = k + 1, and if k <C _max , proceed to (I), if k = C _max , select k with the smallest “h _k ^” as C, and the cache server at that time Output base location as a solution and finish.

  Next, “3.7 Performance Evaluation” according to the present invention will be described with reference to FIGS.

  First, “3.7.1 Evaluation Conditions” will be described.

"NW topology"
In this example, the topology of 36 commercial ISP backbone NWs disclosed on the CAIDA Web page is used for evaluation. FIG. 2 shows the distribution state of the average node order “E (d)” and the maximum node order “Max (d)” of these 36 NWs. From FIG. 2, 36 NWs are classified into the following three groups: “Full mesh”, “H & S”, and “Ladder”.

  “Full mesh”: A topology in which each node is connected to all nodes, and is defined as an NW that satisfies “E (d) ≧ Max (d) −1” and “E (d) ≧ 3”. In FIG. 2, two NWs correspond.

  “H & S”: A node with a high degree (hub node) exists and the maximum node degree Max (d) is large. It is known as a hub-and-spoke type, and it is known that the topology of the air route is this form. Since many nodes can be reached via the hub node, the flow hop length is short, but the load tends to concentrate on the hub node. Here, NWs other than “Full mesh” are defined as NWs satisfying “Max (d) ≧ 10”. In FIG. 2, 12 NWs correspond.

  “Ladder”: A high-order node does not exist, and a topology structure is formed by combining loops. Both the average node order E (d) and the maximum node order Max (d) are small. While the total link length can be suppressed, in order to reach a distant node, it is necessary to go through a large number of nodes, and the hop length of the flow is long. It is known that the highway network is in this form. Here, NWs other than “Full mesh” are defined as NWs satisfying “Max (d) <10”. In FIG. 2, 22 NWs correspond.

"content"
In this example, at each node, to generate a delivery request of content in a Poisson process, it is proportional to the ratio r _n of the occurrence rate population of the city where each node exists. That is, when the request generation rate in the entire NW is “Λ”, the generation rate at the node n is r _n Λ. The number of contents M is “M = 1000”.

Each distribution request generated at node n independently selects content m with probability “q _n (m)”, and this probability q _n (m) is the same for all nodes, and parameter “0.271” It is given as a Zip distribution. Further, the popularity ranking of contents is fixed, and the time (transfer time) _TL required to distribute each content is fixed.

In the simulation, Λ = 1 / second and T _L = 1200 seconds were set. However, the average hop length h of the distribution flow, which is the design of the server base and the evaluation index obtained by the present invention, is not affected by the “request generation rate Λ” and the “transfer time T _L ”.

  Next, “3.7.2 Content diffusion characteristics” will be described.

  All contents are present only in the original server base in the initial state, but are spread to C cache server bases on the NW as the contents are distributed to the user.

  Here, the degree of diffusion of content is evaluated using simulation. However, for the topology of “Cable & Wireless”, when the number of installed cache server bases is set to “C = 5” and the optimal cache server base arrangement that minimizes the estimated average hop length “h ^” is searched brute force ( Hereinafter, the server base arrangement of the Optimum method) is targeted.

  FIG. 3 shows the topology of this NW and the population of each node. In FIG. 3, the top three nodes having the largest population are indicated by stars.

  The average distance to a large population node (ID: 16, 4, 17) is close, and the original server base is located at a high degree node (ID: 10) (in FIG. 3, it is surrounded by a double circle). Shown).

  In addition, cache server bases are arranged in five nodes (ID: 3, 4, 12, 16, 17) surrounded by circles in FIG.

  In this way, the cache server base is a node having a large population (ID: 3, 4, 16, 17) or a node (ID: 12) that has a high concentration of nodes and a high hop length reduction effect in the distribution flow. Can be confirmed.

  4A, five cache server locations (ID: 3, 4, 12, 16 in FIG. 3) when the total cache server capacity (cache server capacity of the entire NW) K is “K = 1000”. , 17) shows the time series of the number of contents cached.

  Cache server bases installed on nodes with a large population receive more distribution requests, so the number of content caches increases rapidly. Further, since the capacity B of each cache server base is “B = 200”, it becomes constant after the number of caches reaches “200”.

  FIG. 4B shows a time series of the total number of copies existing on the NW including the original server for each of the contents having the popularity ranks of the first, tenth, 100th, 200th and 1000th. .

  From the contents shown in FIG. 4 (b), it is found that popular content exists in all cache servers for almost all time, and low popularity content exists only in the original server for all time. It can be confirmed that the number of copies of content in the vicinity of the cache server capacity B (B = 200) varies.

Further, in FIG. 4 (c) the number of copies the contents at the time after the start of the simulation ¹⁰⁶ seconds have elapsed by plotting the popular descending. From the content shown in FIG. 4 (c), it is confirmed that popular content exists in many cache servers, and that there is a large variation in the number of copies of content whose popularity ranking is near capacity B (= 200). it can.

  As described above, as a result of updating the caches of all the cache server bases through which the flow passes using the LFU, the number of copies of the content whose popularity ranking is near the capacity B has a large stochastic fluctuation, but the higher B It is expected that the estimation of the average hop length according to the above equation 7 assuming the existence of individual contents is sufficiently accurate.

  In FIG. 4D, the normalized average hop length (NAH: Normalized Average Hop count) is the cache server base for the four cases of the total cache server capacity K (K = 10, 100, 1000, 10000). The result of plotting with the number C varied is shown.

  However, in FIG. 4D, NAH is a value obtained by dividing the average hop length of the distribution flow by the average hop length when all contents are distributed from the original server without using any cache server. In addition, the actual measurement values obtained by the simulation are plotted with points, and the values calculated from the estimated values obtained by the above equation (7) are indicated by solid lines.

  In either case, C (number of cache server bases) was given and the cache server bases were arranged by the Optimum method. When C is large, since the variation in the number of copies of content whose popularity ranking is in the vicinity of the capacity B becomes large, the estimation accuracy is slightly deteriorated.

  However, it can be confirmed that the estimated value of the average hop length obtained by the above formula 7 is sufficiently accurate in a wide area where the cache server capacity K and the number of cache server bases C of the entire NW are large.

  Therefore, various designs and NAH are evaluated by using estimated values without using simulation.

  Hereinafter, “3.7.3 Comparison of server base arrangement methods in Cable & Wireless” will be described.

  Here, the effect when the cache server location method based on the centrality of the present invention (hereinafter referred to as the Centrality method) is applied to the “Cable & Wireless” topology.

  For comparison, the results of simply placing nodes in descending order of the node order (hereinafter referred to as “Degree method”) and arranging sequentially from nodes having the largest population (hereinafter referred to as “Population method”) are also shown.

  Since many flows go through high-order nodes and many distribution requests are generated from nodes with a large population, any of the comparison methods is expected to reduce the average hop length.

  FIG. 5A shows the normalized average hop length NAH when the cache server capacity K of the entire NW is changed. The NAH decreases with an increase in K (total cache server capacity), and the amount of traffic flowing through each link is reduced. However, the NAH reduction effect associated with an increase in K is significant in a small K region.

  When the cache server base is arranged based on the order, the performance is greatly deteriorated as compared with other design methods. However, the Centrality method and the Population method achieve performance substantially equivalent to the exact optimal solution (Optimum method).

  In addition, when the cache server base position in each placement method was confirmed by changing the cache server capacity K of the entire NW in the range of 10 to 10,000, the Population method is slightly different from the Optimum method, but the Centrality method is It was confirmed to be completely consistent with the Optiumum method.

  FIG. 5B shows a state in which the optimal number of installed cache server bases C is similarly plotted. As shown in FIG. 5B, only the Degree method protrudes and the optimal number of installed cache server bases C is large. However, in the other three design methods, approximately the same number of C is optimal. The optimum C increases with an increase in K (total cache server capacity), but in an area where K is small, the optimum value of C is small except for the Degree method.

  In order to investigate this reason, FIG. 5C shows the NAH (normalized average hop length) when the optimal number of cache server bases C is fixed and the capacity B of each cache server is changed. It can be confirmed that the NAH reduction effect associated with an increase in B is significant in a small B region.

  For this reason, when K (total cache server capacity) is small, rather than increasing the number C of cache server bases, it is effective in reducing NAH to suppress C and increase the capacity B at each base instead. The optimum value of C is small.

  Next, “3.8 Comparison of Arrangement Methods in Various NWs” will be described.

  Up to this point, evaluation has been performed for “Cable & Wireless network”, but the effects of the cache server location method will be compared below for various NWs.

  First, three approximate optimal design methods are compared with the Optimum method. However, since it is difficult to obtain a strict optimal solution for an NW with a large number of nodes N, normalization is performed for 18 NWs with N being 22 or less. Compare the average hop length NAH.

  6A to 6C, NAH (normalized average hop length) in each of the three approximate optimum design methods when the cache server capacity K of the entire NW is “K = 100” is the Optimum method. The results plotted with respect to NAH are shown. In FIGS. 6D to 6F, the results when K = 1000 are similarly shown.

  In each of FIGS. 6A to 6F, when plotted on the straight line y = x, it means that the approximate optimum design method achieves the same performance as the exact optimum solution. ), (C), (d), (f), it can be confirmed that using the Centrality method or Population method, exact optimal solutions can be obtained in all 18 NWs. ) And (d), when the Degree method is used, the performance is greatly deteriorated.

  7A to 7C show the results of comparing the performances of the Centrality method and the Population method for all 36 NWs. In FIGS. 7A to 7C, scatter plots of NAH (normalized average hop length) in the Centrality method and the Population method are used for 36 NWs, and K (the cache server capacity of the entire NW) is three. Two cases (K = 100, 1000, 10000) are shown respectively.

  From the contents of FIGS. 7A to 7C, it can be confirmed that the performance of the population method is deteriorated as compared with the centrality method in many large-scale NWs in which the total node set N exceeds 22. Further, it can be confirmed that the degree of deterioration increases as the total cache server capacity K increases and the number C of cache server bases increases.

  For example, as shown in FIG. 7B, when K = 1000, the population method deteriorates the NAH by about 10% to 20% compared to the centrality method. When both the total node set N and the number C of cache server bases are small, sufficient performance can be obtained even if the cache server bases are arranged in order from the node with the largest population, but N (all node sets) and C (cache server bases) Number) is large and the installation pattern of the distribution flow is complicated, it is effective to use the Centrality method shown in this example.

The computational complexity of the three approximate optimal placement methods (Population method, Degree method, Centrality method) is O (N ³ ) (N: all nodes set), and there is no positive reason to use the Population method or Degree method. In any case, it is desirable to use the Centrality method.

  Further, from FIG. 7B, the carrier server CDN has a limited amount of cache server capacity on the NW, which is about the same as the original server capacity (K (total cache server capacity) = M (total number of contents) = 1000). It can be confirmed that the average hop length of the distribution flow can be reduced by about 30% to 70% by appropriately arranging at the base.

  This means that the total traffic amount flowing through the NW and the average traffic amount of each link are similarly reduced by about 30% to 70%.

  Also, the Ladder type NW, which tends to have a long hop length between nodes, has a higher effect of the carrier type CDN than the H & S type NW.

  8A to 8C, the normalized number of cache server bases (C / (N-1)) when K is optimally arranged using the Centrality method is K (total cache server capacity). The results plotted with respect to are shown separately for each NW type (Full mesh networks, H & S networks, Ladder neoworks).

  The Ladder type NW shown in FIG. 8C has a high effect of reducing the average hop length of the flow by installing the cache server base, and it seems that it is effective to install it at many bases even if the capacity B is reduced. It is done.

  For example, when the total cache server capacity K is about M (total number of contents), in many NWs, the optimal number of cache server bases is about 10% to 40% of “N (all node set) -1”. It is desirable to install a large-capacity cache server base limited to a limited location.

  As described above with reference to FIGS. 1 to 9, in this example, in the carrier type CDN, the original server base and the cache are minimized so that the average hop length of the distribution flow is minimized and the average load of each link is minimized. Optimize server locations.

  In other words, the carrier builds a large-capacity data center at a limited base in its network (NW), and in the carrier-type CDN that operates the CDN, one original server base that always accommodates all content is arranged. Further, based on the distribution request from the user, for example, cache server bases that are updated by LFU are arranged in a plurality of data centers.

  For example, when an NW topology and a frequency of distribution requests for each content are given from each node, one original server base and an arbitrary number of cache server bases are set so that the average hop length of the distribution flow is minimized. Optimal placement.

In particular, for each node n, the centrality H _o (n) of the node n is defined by the reciprocal of the sum total of the number of hops on the shortest path from the node n to each node, which is weighted by the request occurrence ratio. Place the original server on the node with the highest degree.

  Further, the cache server capacity K of the entire NW is given, and the capacity B of each cache server is given as a fixed value (B = floor (K / C)) with respect to the number C of the arranged cache server bases. Including the optimal placement of cache server bases.

In addition, h _{n, min} and h _{n, o} are respectively the hop distance from the node n to the shortest hop distance among the S server bases including the original server base, and the original server to the node n. a hop distance, further, the distribution request of Q n _(m) is from the node n is the cumulative distribution function of a probability q _{n (m)} is intended for content m, and the request generation ratio of r _n from node n Then, the average hop length of the distribution flow is estimated by the above equation (7).

Further, when h _{i, min, Nj} is the hop length from the node i to the nearest server base in a state where the cache server base is installed at N _j , _j cache server bases are already in the node set N _j (N ₀ is N \ O excluding the original server base O from the total node set N), and the estimation obtained by installing the cache server base in the node n (nεN \ N _j ) The reduction effect _{En, Nj} of the average hop length h ^ is calculated by the above _equation 8 _, and the cache server bases are arranged in order from the node with the largest _{En, Nj} .

  According to this example, in the carrier type CDN, the original server base and the cache server base can be optimally arranged so as to minimize the average hop length of the distribution flow and minimize the average load of each link. It is possible to minimize the total amount of traffic flowing in the network.

  In addition, this invention is not limited to the example demonstrated using FIGS. 1-9, In the range which does not deviate from the summary, various changes are possible. For example, in this example, the server arrangement in the carrier type CDN is described, but the present invention can be applied to other server arrangements in a centrally managed network.

  That is, in a network that distributes content from a server computer to a requesting client computer via the network, one original server system that stores all the content as the server computer, and this original server system stores A plurality of cache server systems that store some of the contents and distribute the contents in response to a distribution request from the client computer and perform cache control by LFU on the distributed contents are provided, and the original server system and the cache server system It is good also as a structure which each arrange | positions to the node selected by this invention.

  The computer configuration example of this example may also be a computer configuration without a keyboard or optical disk drive. In this example, an optical disk is used as a recording medium. However, an FD (Flexible Disk) or the like may be used as a recording medium. As for the program installation, the program may be downloaded and installed via a network via a communication device.

It is a block diagram which shows the network structural example of the carrier type CDN which provided the server arrangement | positioning system which concerns on this invention. It is explanatory drawing which shows the example of a dispersion | distribution situation of the average node degree of 36 networks, and a maximum node degree. It is explanatory drawing which shows the topology example and node population situation example of Caboe & Wireless. It is explanatory drawing which shows the example of a simulation result of this invention with respect to Caboe & Wireless. It is explanatory drawing which shows the example of a design result by this invention with respect to Caboe & Wireless. It is explanatory drawing which shows the performance comparative example in 18 networks. It is explanatory drawing which shows the performance comparative example in 36 networks. It is explanatory drawing which shows the number of normalization cache servers in 36 networks. It is a flowchart which shows the server arrangement | positioning procedure example which concerns on this invention by the server base arrangement | positioning apparatus in FIG.

Explanation of symbols

  101: Server location arrangement device, 102: Central control station, 103: Content server, 104: Server selection device, 105: User terminal.

Claims (10)

A server placement method for placing a plurality of server computer systems for delivering content to client computers via a network by programmed computer processing,
As a process execution procedure of the programmed computer,
A first procedure of selecting one node from each node on the network and arranging an original server system for storing all contents in the selected node;
A second procedure for selecting a plurality of nodes from each node on the network and arranging a cache server system for caching and distributing contents stored in the original server system to each selected node. Server placement method. The server placement method according to claim 1,
In the first and second procedures,
Enter the network topology information and the frequency of occurrence of distribution requests for each content from all the above nodes.
A server arrangement method comprising selecting a node on which the original server system and the cache server system are arranged so that an average hop length of a distribution flow is minimized. The server placement method according to claim 2,
In the first procedure,
For each node n on the network, the centrality of the node n is the reciprocal of the sum total of the number of hops on the shortest route from the node n to each other node, weighted by the frequency ratio of the distribution requests. Define H _o (n)
A server placement method, wherein a node having the maximum centrality H _o (n) is selected and the original server is placed. The server arrangement method according to claim 2 or 3, wherein:
In the second procedure,
Each cache server system uses a floor function formula “B = floor (K / C)” with a fixed value of the capacity B of each cache server system and the capacity K of the cache server system of the entire network and the number C of each cache server system. A server placement method comprising selecting a node on which each cache server system is placed, including the number of server systems C. A server arrangement method according to any one of claims 2 to 4,
In the first and second procedures,
Among the S nodes in which the original server system and the cache server system are arranged, the hop distance h _{n, min} to the node having the shortest hop distance from the node _n, and the node from which the original server system is arranged to the node a cumulative distribution function Q _n (m) of the hop distance h _{n, o up to} n and the probability q _n (m) that the distribution request from the node n is for the content m, and the distribution request from the node n by using the generation ratio r _n, from equation 1 below, the server arrangement wherein the calculating the estimated value of the average hop length of the delivery flow to (h ^).

A server arrangement method according to any one of claims 2 to 4,
The second procedure is as follows:
Hop length from the node i to the closest node in the set N _{j of} each node in which each of the j cache server systems is arranged (a set N ₀ excluding the node in which the original server system is arranged in j = ₀ ) _{hi, min, Nj} ,
Hop length _{h n} from the node i to the set _{N j} nodes outside the n (n∈N\N _j), by using the _i, from Expression 2 below,
With the j cache server systems already installed in each node included in the set N _j , the estimated value “average hop length of the distribution flow” obtained by newly arranging the cache server system in the node n “ h ^ ”reduction effect _{En, Nj} ,
A server placement method characterized by selecting a node to place a cache server system in order from a node having the largest calculated reduction effect _{En, Nj} .

An arrangement method of a server computer in a carrier type CDN system that distributes contents cached in a server computer by a CDN to a requesting client computer in a network operated by a carrier (carrier type network),
The node in which each of the said original server system and the said cache server system is arrange | positioned is selected by each procedure in the server arrangement | positioning method in any one of Claims 1-6, The said original server system and the selected node are made into the selected node. While arranging each of the above cache server systems,
A server arrangement method in a carrier type CDN, wherein the original server system and the cache server system are configured by clustering a plurality of server computers. A server arrangement system for arranging a plurality of server computer systems for distributing contents to client computers via a network by programmed computer processing,
As a means of performing programmed computer processing,
Comprising means for executing each procedure in the server placement method according to any one of claims 1 to 7,
By this means, selection of a node on which each of the original server system and the cache server system is selected, and arrangement of the original server system and the cache server system on the selected node are performed. system. A carrier type CDN system that distributes content cached in a server computer to a requesting client computer in a network operated by a carrier (carrier type network),
As a device that executes programmed computer processing,
A means for executing each procedure in the server placement method according to any one of claims 1 to 7, comprising: selecting a node on which each of the original server system and the cache server system is placed; A server base location device for placing each of the original server system and the cache server system on the selected node;
A centralized control device that receives content distribution requests from all client computers and sends out instructions to select a cache server system suitable for content distribution to the client computers;
Upon receiving the selection instruction sent from the centralized control apparatus, a cache server system suitable for the distribution of the content to the client computer is selected, and the distribution of the content to the client computer to the selected cache server system A carrier-type CDN system, characterized in that a server selection device for instructing is provided.   The program for making a computer perform each procedure in the server arrangement | positioning method in any one of Claims 1-7.

Images