Caching in Information-Centric Networking Sen Wang, Jun Bi, Zhaogeng Li, Xu Yang, Jianping Wu Network Research Center, Tsinghua University {wangsen, lizhaogeng, junbi, yangxu}@netarchlab.tsinghua.edu.cn;
[email protected] cache hit instead of simplly couting the hit number as LFU ABSTRACT does. Several series of simulations are conducted over Information-Centric Networking is (ICN) [1] gained simple topology and practical ISP topology to evaluate our increasingly concerns, as an important direction of the analysis and proposed cache policy. We use synthetic Future Internet Architecture research. Although In-network traffic, which is generated from content-based traffic model caching is considered as one of the most significant deriving from traffic study of Web caching to test these properties of ICN, the cache policy for ICN is still little caching policies. explored. In this paper, to explor the properties of the optimal cache policy, we formulate the in-network caching In contrast to other caching approaches for redundant problem of ICN into Mixed-Integer Linear Programming elimination (e.g. Web caching, Packet-level Redundant problem and find that LFU-like cache policies is supposed Traffic Elimination), the caching of ICN disccussed in this to perform well when considering minimizing the average paper is based on explicit identifiers for receivers and accessing hops to get content. We also propose a novel pub/sub communication paradigm of ICN. cache policy named LB (Least Benefit) with taking into The rest of the paper is organized as follows. In Section 2, account the distance factor besides accessing frequency and some basic assumptions are made to make our problem a new forwarding scheme with shallow flooding (FSF for background clear. With Section 3.1 simply state the main short)to improve the performance further,. Our simulation challenges for solving caching problems of the interresults show that with in-networking caching, the average domain setting, this paper focuses on the cache policy hops of the ICN network can be reduced significantly by applicable to intra-domain setting. Subsection 3.2 presents nearly 50% and with some simple improvement such as LB the Linear Programming formulation of the in-network and FSF the average hop can be reduced future. caching problem of ICN. In Section 4, the properties of the Keywords optimal caching are studied and the proposed cache policy Information-Centric Networking; Caching; Future Internet; named LB is derived according to these properties. In Section 5, we evaluate our analysis findings and proposed 1. INTRODUCTION cache policy with simulation based on NS3. We conclude As a Future Internet Architecture proposal, ICN intends to in Section 7. motivate the architectural transition from today’s hostcentric Internet architecture to information-centric in order 2. BASIC ASSUMPTIONS to disseminate efficiently and scalably the enormous and With the aim of being independent of any specific diverse informations generated by a variety of applications. approaches aforementioned, in this paper we extract the In this research area, many approaches have been proposed ICN structuring architectural properties and assumptions as such as PSIRP [12], NetInf [13, 14], PURSUIT [10], CCN the following. We refer to the content identifier as CID [15], DONA [16] and NDN [11]. In-network caching is (content ID) which names a content (e.g. a video or a considered as one of the most significant properties of ICN, named single packet) uniquely. The content transimssion which can be leveraged to improve network performance mechanism adopts the Pub/Sub communication paradigm significantly in terms of improving utilization and reducing which is embeded in most of the approaches propagation delay. aforementioned. It means, a request is sent to get a specific content with the CID specified in its packet. The request is So far, the cache policy for ICN is still little explored. In routed to the original content without awareness of any [2], preliminary evaluation on network performance cached copy in routers by some routing mechanism (e.g. improvement by random autonomous caching is exhibited OSPF). We make this assumption because of considering with very simple topology and scenario. In this paper, we that knowing current caching state of the whole network consider the benefits of deploying of ICN and inwould imposes the routing system significant burden in networking caching to the ISP network and study the terms of maintaining extra states. Along the path of a impact of employing different caching policies to the request, any router caching a copy of the desired content overall network performance of ICN. In this paper, the could respond to this request with its copy. We assume the optimal cache for ICN is explored. We formulate the incontent will go back to the requester along the same path network caching problem of ICN into Linear Programming which means an accurate symmetric routing. While problem (Integer Programming, more accurately) and forwarding a content, an intermediate router can choose to analyse the properties the optimal cache policy should cache the content according to its own caching policy. comply with. A novel cache policy refered as LB (least benefit) is proposed which takes into count the benefit of a
3. CACHING IN ICN 3.1 Inter-domain Caching In inter-domain category, the inter-domain topology of Internet is determined mainly by the commercial relationship between ISPs rather than the network architecture itself. Accordingly when we tall about the inter-domain cache policies of ICN, the commercial relationship between ISPs should be taken into account. So far, all the proposals of ICN assume the same relationship between ISPs as that of the current Internet. Actually through analysis, [10] argues that i) different roles in the current valley-free commercial relationship model may view in-network caching differently according to the impact on its revenue; ii) the richer network service model of ICN opens up the possibility of using more efficient policies, leading to better network utilization and lower costs; iii) In ICN, the accouting model of the peering balance changes from the current model. These there findings implicate that preceding the study of inter-domain caching policies of ICN, the evolution of the commercial relationship between ISPs should be studied first. So far, how to make the ICN design policy-compliant and incentive compatible is still an open issue. This paper foucses on the intra-domain caching policies for ISP which can benefit the deployed ISP immediately.
3.2 Intra-domain Caching For ISPs with ICN deployed in its network, the main concern in our opinion is how to cache the content generated from its customers or imported from exterior network optimally so that the network resource consumed by the transfer of these content or/and the transmission latency was minimized. Considering the static case, the innetwork caching problem can be described as the following. Given an ICN network, the request rate from each router to each content, the storage capacitiy of each router, a set of initial contents, an assignment of resident routers of these contents and a fixed routing path from each router to each content, the goal is to find a feasible assignment of caching copies of each content to routers in order to minimize the overall resource comsumption of the network. Now we start to formulate this problem into a MILP problem as follows.
is the ratio between the number of requests initiated from node v for content c and the number of all the requests. Obviously, we have ∑ ∑ . Routing path: We assume that there is only one routing path between any pair of nodes. denotes the single path between node u ) gives the length of this and node v and the function ( path. The k-th node in the path from node u is denoted by the return value of the function ( ) for any ( ). An assignment of in-network caching: We use a series of binary variables * + to describe the caching state. If node v has cached the content c in its storage, the has the value 1, otherwise 0. For a request, the network resource consumed is estimated as the product of the number of the hops traversed to get the first copy of the desired content and the size of corresponding content. Then the overall objective fuction is as follows.
Traffic demand:
(
( (
)
(
(
())
(
( ))
( ) ))
(
(
(
( )
)
( ) )
( )
( ) )
(
( ) ))
is
the
minimal number of hops from v to any copy of content c along the routing path ( ) , including the original content resided in node ( ). The objective function (1) is subject to the storage capacity constraint of any node v: ( )
∑
( )
( )
Note that the min function is not a linear function, so we involve a series of additional continuous variables and binary variables for every node v and every content c to linearize the objective function. The binary variables * + indicate whether the is the ( ( ) ) minimal hop number for any i subject to ( ( ) ). The represents the original content. These ( ) ( ) variables are subject to the following constraint: (
Graph construction: The ICN network is represented as an undirected graph ( ); V is the set of nodes in the network; E is the set of edges in the network; is a function. ( ) denotes the storage capacity of the node v; C is the set of initial contents. A content is denoted by a tuple (l,m), where l is the resident node of the content and m is the size. For any content , function ( ) and ( ) return the resident node and the size respectively.
( )
∑∑
( ))
∑
( )
Besides, we have the following constraint for the variables : (
(
( ) )
(
(
(
( ) )
(
( ) )(
)
( ) ))
(
( ))
( ) )
(
( ))
(
(
( ) ))
(
(
( ))
)
( ) Inequality (4) indicates that when has the value 1, the can not have the value 0, which ensure that ( ( ) ) when a node v is indicated to be the minimal, it must have the copy. Inequality (5) associated with (6) assures that is at least larger than one of for any i subject ( ( ) ) to ( ( ) ) or the path length ( ( ) ) and the minimization objective function can naturally ensure take on the minimal value. Finally, we have the ultimate objective function: ( )
∑∑
( )
It is subject to the constraint (2)(3)(4)(5)(6).
4. CACHING POLICY 4.1 Theoretical Analysis To explore the optimal cache policy for the optimization objective function (7), we first study the case in which there is only one requester in the network. In Figure 1(a), node a is the only requester. According to the hypothesis of our model that there is only one path for every pair of nodes in the network, the paths from a node to any other node form a tree rooted at node a as Figure 1(a) shows. For analytical brevity, we further assume each content has the same size cs and the storage capacity of any node v is interger times of the content size cs, namely . Let d denote the maximum hop number from node a to any other node. We partition the set of all the nodes into d subset according to the node’s distance to node a, namely . represents the set of nodes whose distance away from node a is i. g c
Theorem 1: If an assignment A is among the optimal cache assignments, it must have the property that any content cached in node n, its request rate is not smaller than that of any content cached in nodes which reside in the subtree rooted at node n. Proof: For contradiction, suppose such an optimal cache assignment exists so that there are at least one pair of contents does not comply with that the request rate of one content cached by root node of some subtree is smaller than that of another content cached by the node within the subtree. In this assignment denoted by A, suppose one of these pairs of content is d and and they reside in the node n(d) and ( ). Their hop numbers from node a are ( ) and ( ) respectively. For ( ) is one of nodes within the subtree rooted at n(d), ( ) is larger than ( ) . According to hypothesis, the request rate of content , , is larger than that of content d, . For content is cached in node ( ) within the assignment, the resident node of the original content is in the subtree rooted at ( ) according to the rational caching assignment definition. Then we can change the caching node of the pair of contents and end up with a new caching assignment. The new assignment will get additional profit ( ) ( ) and profit penalty smaller than ( ( ) ) (smaller in the case that the hop number of the resident node of the original conten from node a is larger than ( )). So the final additional profit will be larger than ( )) and the new assignment is ( ) ( ( ) found a more optimal one which is contradicted with that the former assignment A is a optimal one. The Theorem 1 is proved. ■ Let Sub(n) denotes the set of nodes within the subtree rooted at n. To generate a caching assignment complying with Theorem 1, we conceive a greedy algorithm as follows:
f
Algorithm 1: Greedy Caching
e
b
content cached in node n, its request rate is not smaller than that of any content cached in nodes which reside in the subtree rooted at node n.
1:
;
2:for(i =0; i
