International Conference on Computing, Communication and Automation (ICCCA 2016)

A Constructive Review of In-Network Caching: A Core Functionality of ICN Anshuman Kalla, Sudhir Kumar Sharma Department of Electronics & Communication Engineering Jaipur National University, India Email: {kallajnu,sudhir.732000}@gmail.com

Abstract—Over the years, caching has been leveraged (and established) as an add-on functionality to enhance network performance. However, Information Centric Networking (ICN) conceives caching at network layer (i.e. beyond the premise of end-to-end principle) thereby making it one of the core functionalities. Further, ICN advocates named-content (another core functionality) that allows content-consciousness within networks. Together the two functionalities result in content-aware ubiquitous in-network caching that has received significant attention of researcher all-around. The aim of this paper is to probe in depth and review some of the work done pertaining to in-network caching in ICN, in order to understand their aims, assumptions, approaches, simulation set-ups used, network topologies exploited, traffic pattern fed-in, performance metrics used, parameter(s) tuned-in to optimize the performance and the significance of the results obtained by the researchers. The paper also lists out advantages of in-network caching, related issues, factors that affect in-network caching and relevant performance metrics. The paper intends to assist researchers who are searching ways to put-forward (an acceptable) proof of their ideas.

I. I NTRODUCTION Early 1970s marks the ear of design and development of todays networking ecosystem. Efficient sharing of expensive resources was the ultimate aim then. Since then numerous technological advancements (like proliferation of economical hand-held networking devices, multiple simultaneous connectivities, availability of high speed data communication links, advancements in multi-core processors technology, consistent decline in the cost of memory for data storage etc.) have fallen in line to support the flawless evolution of networking facility. Unfortunately, in spite of years of maturity and all related technological developments, it is being experienced that networking still falls short of users’ expectations. Some of the issues ( [1], [2]) that have in a way plagued current TCP/IP networking architecture over the years are: •

Data Dissemination & Service Access: Today Internet is prominently being used for dissemination/retrieval of

c

2016 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all the other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works c ISBN: 978-1-5090-1666-2/16/$31.00 2016 IEEE

•

•

•

•

•

staggering amount of content and service access, for which it was not tailored. Named Host: TCP/IP networking translates (according to DNS mapping) name of every content (to be retrieved) to an IP address which signifies the location of a host (serving that content) in network space. Thus Internet today comprises of named-hosts instead of named-contents. Mobility: Mobility causes intermittent connectivity which might lead to change in IP address thereby compelling ongoing TCP/IP application(s) to restart. Availability: It implies existing content or service should be always available to the users with high reliability (preferably with low latency). Security: So far, security and authenticity of data has been realized at network-level by ensuring communication over a secured channel with a trusted server. Datalevel security is however still missing. Flash Crowd: Leads to network congestion, Denial-ofService (DoS), poor QoS etc. Existing content-oblivious networking is natively incapable to tackle such a practical phenomenon.

To handle these issues, the trend being followed so far is to progressively design and deploy a fix for every issue that is encountered. To mention some are CDN (Content Distribution Network) and P2P applications for efficient data dissemination, DNS for resolving an identifier (URL) to a locator (IP address), MobileIP and Shim layer for supporting mobility, DNSSec & IPSec for ensuring security, web caching & intelligent DNS for availability issue etc. As a result of such patches the networking today is surviving critically. In order to permanently conciliate the problem in unified manner, numerous clean slate Information Centric Networking (ICN) approaches have been recently proposed. Albeit, the design details of the proposed architectures are different but they unanimously talks about retiring the host-centric (host-tohost) communication model and bring in place the contentcentric (host-to-content) communication model. Readers are referred to seminal works [1], [2], [3], [4], [5], [6], [7], [8] for CCN, NDN Project, 4Ward Project, NetInf, PSIRP, PURSUIT, DONA and SAIL respectively. Next section II introduces in-network caching, its advantages, issues involved in cache management, factors that affect

International Conference on Computing, Communication and Automation (ICCCA 2016) caching and metrics to evaluate its performance. Some of the techniques proposed so far in context of on-path, off-path and hybrid caching have been exhaustively reviewed in sections III, IV and V respectively. Section VI presents conclusion. II. I N -N ETWORK C ACHING IN ICN Content-aware pervasive in-network caching is configured as one of the core functionalities of ICN. It entails that every node, in addition to routing, buffering and forwarding operations, should perform caching of (traversing) contents in its Content Store (CS). There is subtle difference between content store (commonly called as cache) in CCN node and buffer in an IP router. Buffer holds an incoming IP packet as long it does not get a clear path to reach next hop closer to the destination node. The buffer memory is recycled immediately after the packet is forwarded to next hop. This happens because network is deprived of intelligence to know what is inside the IP packet. Whereas in CCN due to named-content, the network is aware of what is flowing through it. Moreover, IP buffering seeks higher eviction rate where as caching in CCN lays emphasis on higher retention rate so as to achieve higher cache hit rate. Caching techniques studies so far could be broadly classified as (i) On-path caching, (ii) Off-path caching and (iii) Edge caching and (iv) Hybrid caching. On-Path Caching aims to cache the retrieved contents at the intermediate nodes on the (symmetrical) way back from server to the requester. Accordingly, an interest packet (being forwarded towards the server) taps on-path caches in an opportunistic way. In default on-path caching, a retrieved content is being placed at all the nodes encountered on the path. Thus it achieves reduction in server load and retrieval latency at the cost of high degree of network-level content replication. Redundant content caching together with limited size of expensive content storage results in under utilization of in-network caching capacity. The situation could be further aggravated by multi-path routing which is another inherent feature of CCN paradigm. Section III presents some of the works ( [37]–[57]) carried-out in context of on-path caching. Off-Path Caching could place a retrieved content at any node within a network, without any contrived correlation with the nodes that fall on the path being followed by interest packets to reach the server (in case of on-path caching). That is for every content to be cached, a node is appointed as designated off-path-cache which acts as secondary point-of-source for that content. A thought from the implementation point-of-view would reveal that there is cost to be paid for deploying offpath caching. A part of the cost comes into the picture due to the necessity of deploying an additional deflection mechanism (centralized or distributed) which is required to re-direct the requests, off-the-route, towards the designated off-path-cache. In other words, cache-(content)-aware routing needs to be rolled out. Yet, remaining part of the cost is ascribed to the coordination overheads required among the nodes to resolve content placement. Section IV revisits some of the works [58]– [60] and [62]–[66] done pertaining to off-path caching.

Edge Caching is entirely different then two caching discussed above and opposes pervasive in-network. In edge caching only the nodes at the boundary of a network are enabled with caching capability. The work done pertaining to edge caching [42], [48]–[51] is discussed in section III along with on-path caching. Hybrid Caching tends to profitably couple different caching styles. As an example on-path and off-path caching have been clubbed by researchers. In such approaches an interest packet arriving at ingress node is asked to march towards a server in the same way as it would have moved in case of on-path caching. Additionally at intermediate nodes, based on some information, decision could be taken to sent out (off-the-path) clones of interest packets to explore near-by caches. Such hybrid caching could be viewed as on-path caching blended with opportunistic off-path caching. Section V reviews some of the work [67]–[72] related to hybrid caching. A. Advantages of In-Network Caching in ICN In-network caching in ICN allows re-utilization of cached contents and results in following advantages: Cost Effective Data Retrieval: In-network caching minimizes delegation of traffic concerning the cached contents over egress links thereby minimizes usage of notably expensive transit links/networks and mitigates significant percentage of total traffic approaching the origin server. The merit is well supported by [36] where it is pointed out that inter-ISP links are costlier and are bottleneck. Reduction in Latency: With the contents cached at intermediate nodes which are comparatively closer to requesters, innetwork caching results in effective reduction in latency thus improves Quality-of-Service (QoS) perceived by users. Heavy Load Handling: Under heavy load situation, particularly during flash crowd when there is viral dissemination of an extremely popular content(s), in-network caching potentially transforms nodes into legitimate proxies of origin server thereby inherently tackles heavy load. Efficient Retransmissions: Network wide caching ensures better resiliency to packet losses by quick retransmission from the closest intermediate node where un-corrupted cached copy (of content to be retrieved) is available. Higher Availability: Network caching substantially improves content availability thereby reduces the probability of Denial of Service (DoS) attack to a great extent. Buoyancy to Intermittent Connectivity: Caching also plays a vital role in building inherent resiliency to intermittent connectivity and also allows mobile nodes to act as a network medium for areas uncovered by network. B. Issues Related to In-Network Caching in ICN To the best of our knowledge, there are following broad issues have been worked upon by researcher so far in context to in-network caching in ICN: 1 Cache Placement or Allocation: Where to place the caches (i.e. content stores)? The issue deals with the decision regarding caching facility to be made available at all or

International Conference on Computing, Communication and Automation (ICCCA 2016) selected points in a network. ICN doctrine suggests caches mance. However, popularity estimation is in itself a chalto be placed throughout network whereas some researchers lenging task as it mostly requires tracking of number of [42], [48], [49], [50], [51] advocate to place the caches only requests per content at all or selective nodes. Further the at the edges. Yet some approaches [40], [42] intent to devise monitored number of requests is compared either with a ways to find few best locations within a network where the threshold value (dynamic or periodically adjusted) or with caches should be placed. request statistics of other contents. So far popularity has 2 Cache Size Dimensioning: What should be the size of been implicitly introduced by using real traffic traces ( caches? It deals with the decision of allowing homogeneous [43], [46], [50], [52], [63]) or explicitly introduced by using [38], [41], [43], [49], [52], [59], [62], [64] or heterogeneous Zipf distribution or Zipf-Mandelbrot distribution to generate [40], [42], [48], [50] caches. Moreover, in case of heterogesynthetic workload ( [37]–[42], [44], [45], [47]–[49], [51]– neous caches, it needs to be decided where the cache size [59], [62], [64]–[72]). Several works [38], [44], [45], [46], to be boosted comparatively. In this context, an approach [47], [49], [51], [52], [54] have used content popularity as that could run off-line over a given network to produce a a key factor to actuate the proposed caching schemes. blueprint of cache dimensioning is highly desirable. 4 Popularity Dynamics: Percentage and/or frequency of 3 Content Placement: Where to cache a retrieved content change in popularity of contents represent popularity dywithin a network? For a content to be cached the issue namics. Impact of popularity dynamics on cache perfortries to decide the caches in a network where if placed will mance is worth exploring and was studied in [45]. improve the performance of in-network caching. It could be 5 Latency: The latency involved in retrieving content in terms performed in centralized [66] or decentralized [41], [44], of hop-count or distance has also been used to trigger [47]–[49], [52], [58], [59], [60], [62]–[66], [69]–[71] way caching decision [43], [47], [48], [58], [63], [65], [69], [70]. along with (explicit or implicit) coordination of nodes [43], 6 Bandwidth: Another factor that has been tapped to perform [50], [52], [58], [60], [64], [66], [67]. caching is bandwidth available over retrieval path [54], [55]. 4 Content Selection: What to cache out of huge flow of 7 Cache Size: Yet another factor that heavily influences contents? It involves ways [44], [53]–[57], [66], [69], [72] in-network caching is cache size. A network could be to identify contents that are valuable and should be cached comprised of homogeneous ( [38], [41], [43], [49], [52], preferably. It might appear that content placement in a way [59], [62], [64], [65]) or heterogeneous ( [40], [42], [48], results in implicit content selection from a node’s viewpoint. [50]) sized caches. Cache size has been also used to analyze That is a node caches the selected contents which is outcome caching performance [48], [51], [65]. of content placement. However, it is worth noting that 8 Granularity of Content Caching: It defines the smallest content selection could still be performed after content cacheable unit which could be (i) entire object, (ii) chunk placement decision, particularly when the later is oblivious (block) [44] or (iii) packet. of content’s utility characteristics. 9 Size of Content: To study the effect of content size 5 Replacement policy: Which cached-content should be on performance small-sized and large-sized contents [46] evicted to accommodate an incoming content? That is when as well as heterogeneous sized contents [57] have been an incoming content is to be cached and if content store is leveraged. full then issue regarding content eviction is being taken care 10 Pricing: Financial cost involved in fetching contents from by replacement policy [38], [39], [43], [45], [47], [54], [55]. expensive external links has also been recently tuned-in to prioritize caching of costlier contents [72]. Particularly, for ICN first two issues is to be decided when network is being designed or upgraded. However, last three 11 Mobility: Movement tendency of users has also been geared-in to roll-out pre-fetching based caching [57]. issues; content placement, selection and replacement demand 12 Routing: In general shortest-single-path routing has been a careful consideration from ICN’s point-of-view. considered, however effect of multipath routing on caching C. Factors Affecting In-Network Caching in ICN performance has also been studied by [38]. Next, we enumerate numerous factors that have been studied 13 Spatial Locality: Accessing tendency of users in a particuby researchers. lar geographical area i.e. spatial locality has been considered 1 Network Topology: Number of papers [39], [40], [41], [42], by [51] to accomplish caching. [51], [66] have considered topology awareness as crucial 14 Social Networking: Behavior of socially active users has information for performing caching, whereas [38] states that been also employed to effectuate in-network caching. Concaching performance is altered insignificantly with different tents produced or accessed by these influential users are network topologies. expected to be popular. Authors in [53] thus proposed to 2 Size of Content Population: Represents the total number of cache on priority basis such contents. Network topology was distinct contents for which a network could receive requests. generated using Inet [24] and LastFM [32] & Facebook [33] Size of content population (catalog) plays a vital role in data sets were used to model social relationships between characterizing cache performance [38]. users. Eigenvector and PageRank centrality measures were 3 Popularity Distribution: Popularity profile of contents is employed to discover influential users. evidently an important factor that drives caching perfor-

International Conference on Computing, Communication and Automation (ICCCA 2016) D. Performance Metrics Following are the various metrics used to evaluate performance of in-network caching in ICN: 1 Hit Ratio: The metric denotes number of requests that are successfully satisfied by the contents residing in caches to the total number of requests received. It is one of the prominent performance metrics used by [37]–[39], [40], [45], [49], [53]–[57], [59], [60], [62]–[66], [69], [71]. In general, higher is the cache hit ratio for popular contents better is the overall cache performance. Nevertheless, [72] explained that the same is not always true. 2 Bandwidth Usage: Reduction in the bandwidth usage of external links ( [46], [52], [57]–[59], [69], [72]) or internal links ( [42], [52], [62], [65], [69]) has been considered as an another important metric. At times, link stress, representing volume of traffic passed though the top most heavily loaded links, has also been used [44], [50], [68]. 3 Cache Load: The metric indicates either moderately loaded or over-loaded caches. Moreover, caches could be homogeneously or heterogeneous loaded. Later leads to unbalanced caches [64] that lead to creation of hot spots. 4 Server Load: Reduction in server load has been widely used in [41], [43], [45], [46], [48], [50], [58], [60], [66]– [68], [70] to measure caching performance. 5 Latency: Reduction in content retrieval latency has been considered as another extremely useful parameter to gauge caching performance [37], [38], [40], [41], [43], [44], [46], [48], [50], [51]–[56], [66]–[68], [70], [71]. 6 Cache Diversity: It implies number of unique contents residing in network caches and has been used to assess caching performance [53], [58], [60]. 7 Complexity & Overheads: A caching techniques needs to be simple, light-weight and practically deployable as emphasized by [38], [43], [60], [67], [70]. 8 Fairness: From cache management point-of-view fairness could be in terms of content selection ( [45], [48]) content admission, content replacement, cache partitioning etc. Nevertheless fairness could also be measured in terms of link load fairness [56] and popularity estimation fairness [45]. 9 Resiliency to DoS Attack: Network wide caching establishes nodes as legitimate proxies of origin servers thus empowers them to respond to interest requests with authentic copies of contents. Consequently, caches collectively handles Denial-of-Service (DoS) attack [34]. III. O N -PATH C ACHING D. Rossi and G. Rossini [38] studied the dependency of on-path caching’s performance over numerous factors and pointed out that multipath routing leads to severe cache churn (and high cache replacement error), content-popularity & size of content population are most dominating players whereas network topology affects the minimum. It is also shown that random replacement policy closely follows the performance of any other complex replacement policy. The ratio of router’s cache size to the size of content population was set to 10−5 .

Negligible transmission delay, congestion free regime and infinite capacity of intra-network links were assumed. Authors developed ccnSim simulator [25] over Omnet++. Five real network topologies (Abilene [14], Tiger2 [18], Geant [17], Level3 [12] and DTelekom) and a tree topology were used. As hinted in [38], Gallo et al. [39] explored the feasibility of random replacement policy for network-level caches. In contrast to LRU, random replacement is extremely lightweight and can tackle caching at line speed since it selects randomly, with uniform distribution, content (residing in cache) for eviction. Since the core routers experience more requests from multiple edge routers and witness distorted popularity distribution thus the authors conclude that random replacement policy works well for such routers. LRU, however, is more suitable for edge nodes. D. Rossi and G. Rossini [40] endeavored to perceive the impact of topology conscious heterogeneous cache allocation, over the on-path caching performance, governed by numerous centrality metrics (Degree, Stress, Betweeness, Closeness, Graph and Eccentricity Centrality). The ratio of cache size per node to content population considered was 10−5 . The paper assumed congestion fee environment and internal links to have infinite capacity. Simulator ccnSim [25], real topologies (Abilene [14], Tiger2 [18], Geant [17], Level3 [12] and DTelekom) and Zipf distribution (α ∈ [1.25, 1.5]) were used. Authors concluded that degree centrality turns out to be simple and doable kind of metric for cache size allocation. Chai et al. [41] also leveraged betweenness centrality to improve performance of on-path caching. However, they used it for strategic content placement instead of cache allocation. Authors proposed distributive estimation of betweeness centrality based on ego network since at times network topology might not be stable. Cache size per node considered was 10% of content population. Zipf law (α = 1), real topology from CAIDA [10] and scale free topology based on BA power law model [11] were used. The proposed caching improves hit rate, reduces retrieval delay and off-loads server(s). To some extent sharing the objective with [40], Wang et al. [42] devised topology-driven cache placement and size dimensioning scheme to roll-out on-path caching. Higher cache size is allocated to those nodes which offer better reduction in latency based on their topological positions. Authors used Zipfian popularity distribution (α = 1), shortest path routing, fabricated topologies (based on BA [11] & WS [13] model) and aggregate network caching capacity of 1% of content population. Results indicate that the proposed optimal cache allocation technique performs progressively better with increase in network size. It is also concluded that for small size networks it is edge caching that performs better. Ming et al. [43] coupled age-based caching with implicit cooperation between nodes to develop Age-Based Cooperative (ABC) replacement policy. The rationale is to push the popular content’s caching progressively towards the edge of the network. Every content carries an age which is being updated by every node before caching it based on two parameters; popularity of the content and distance of the current node

International Conference on Computing, Communication and Automation (ICCCA 2016) (willing to cache the content) from the server. Real topology CERNET2 [9] was used. The results exhibited that ABC offers lower average delay and off-loads the server. Cho et al. [44] performed conservative chunk level caching WAVE by selecting the chunks based on content popularity. The selected chunks are cached only at the first hop downstream. The idea is to cache exponentially increasing number of chunks (of a given content) progressively closer to the requester, with the increase in number of requests for that content. Authors used synthetic topology generated by GTIMT [23], Zipf distributed workload (α = 0.85) and network cache capacity of less than 1% of content population. It is shown that WAVE offers reduction in average latency and link stress with increase in hit ratio. While negating the efficient applicability of recency and frequency based cache replacement policies to CCN, Kang et al. [45] tried to categorize the contents and performed categorylevel monitoring for popularity estimation as doing so is relatively simpler and imposes less computational overheads. Popularity dynamics was made to vary every 30 sec. The cache size per node was considered to vary from 10% to 30% and OPNET [29] was the simulator used. Results shows that the proposed Recent Usage Frequency (RUF) progressively outperforms all the other cache replacement policies (in terms of average cache hit ratio and average server load) as the network cache size is reduced. Wang et al. [47] proposed Least Benefit (LB) replacement policy based caching where a parameter named benefit is maintained per content which signifies the worth of retaining a content in a cache. Moreover, to estimate the worth of a cached content, gain offered in terms of hop-count reduction and usage frequency has been used. Over the time, benefit of a cached content is updated (i.e. incremented by hop reduction offered by the cached copy) every time a cache hit occurs. Psaras et al. [48] tried to ameliorate performance of on-path caching by minimizing redundant content replication and by ensuring fair share of cache facility (along the path) among the on-going multiple active flows. The designed ProbCache algorithm allows probabilistic content placement at nodes along a retrieval path. Zipf popularity distribution (α = 0.8) and both heterogeneous & homogeneous caches were used. The work implies that it is better to load edge nodes with more cache capacities. The results demonstrated that ProbeCache technique performs better in terms of latency and server load. Wang et al. [49] improved on-path caching by means of conservative caching of contents, governed by three proposed rules for content placement. First rule implies unconditional caching of retrieved contents at the first hop downstream. Whereas second and third rule signify caching of retrieved contents at remaining intermediate nodes and edge node respectively based on contents’ requests. The paper used Omnet++ [29] simulator and hierarchical tree topology. The workload was generated using ProWGen [35] and requests followed Zipfian popularity distribution (α ∈ [0.74, 0.78, 0.92, 0.96]). The performance gain is significant in case of edge nodes especially when the cache size is small. Thus authors indicate

influential role of edge caching as mentioned in [48]. Fayazbakhsh et al. [50] posed a fundamental question of whether ubiquitous in-network caching and name-based cache-aware routing, needs to be adapted in strict cleanslate manner. Authors aimed to acquire the elegance of ICN without disturbing the legacy Internet. Paper evaluated two versions of on-path caching (with shortest path routing and with nearest replica routing) against two versions of edge caching (with and without cooperation among edge nodes). Real topologies ( [14], [17] and [12]) and CDN as well as synthetic request traces were used. The results suggest that performance gap between on-path and edge caching is more or less acceptable in the view of fact that edge caching can be incrementally deployed over the existing network architecture and estimates less overall expenditure. The study strongly favors edge caching as suggested by [48] [49]. In line with [48], [49], [50], Dabirmoghaddam et al. [51] made comparative study through formal evaluation for optimal caching, edge caching and default on-path caching under different scenarios: (i) hierarchical (tree structured) network topology and random (unstructured) network topology and (ii) request traffic traces with Independent Reference Model (IRM) and with spatio-temporal reference locality. IRM assumes that subsequent requests are absolutely uncorrelated. However, reference locality indicates that present requests are dependent on past requests (i.e. temporal locality) and geographical area of origin (i.e. spatial locality). Interestingly results showed that edge caching consistently performs better than on-path caching. With this conclusion the authors reinforce the question raised by [50]. Li et al. [52] attempted to minimize inter-ISP traffic and average latency by leveraging on-path coordinated caching. Synthetic topology generated by GT-ITM [23] and real topologies UUNet Alternet (AS701), UNINETT (AS224) and University of Wisconsin (AS59) were exploited. Cache size per node was varied from 0.01% to 0.12% of the size of content population. Authors proposed two popularity-guided content placement techniques named as TopDown and AsympOpt. Results showed that AsympOpt works close to optimal in terms of reduction in inter-ISP traffic and latency. Previous works [43], [47], [48], [58], [63] and [65] exploited delay in terms of hop-count or distance ratio while considering congestion free regime. However, the fact is that congestion might result in (i) different delays observed at end user for the paths with same hop-counts and (ii) effectively higher delays for paths with lesser hop-counts as compared to paths with more hop-counts. Thus Badov et al. [54] introduced congestion awareness as yet another dimension to in-network caching. The network caching capacity considered was 5% of size of content population. Shortest path routing and Zipf popularity distribution (α = 0.8) were used. The proposed Congestion Aware Caching (CAC) policy outperforms other caching techniques from viewpoint of average retrieval delay. However, the same is not true for average hit rate metric. Nguyen et al. [55] in a way adds to the work of [54] and pointed out that estimating congestion by considering

International Conference on Computing, Communication and Automation (ICCCA 2016) bandwidth utilization at single bottleneck link over the delivery path might lead to sub-optimal caching in case there are multiple bottleneck links present over the same path. Thus authors take into account the bandwidth utilization pattern over entire retrieval path to compute common congestion price which governs caching decision. Grid topology and real topology from Rocketfuel [12] were exploited. Cache size per node was taken to be 2% of content population. The results showed that with the increase in rate of requests, content delivery time reduces. With somewhat similar objective considered in [54], [55], Carofiglio et al. [56] put-forward yet another variant of latency conscious in-network caching. Probabilistic content selection is effectuated by estimating latency which considers haul distance as well as network congestion. However in contrast to [54], authors performed probabilistic content caching depending upon retrieval latency of a given content compared with latencies of all the contents witnessed so far. CCNPL-Sim [28] simulator, Zipf distribution (α = 1.7) and single cache & tree topologies were used. Results exhibited enhanced user experience, balanced link load and reduced miss probability. Users with ongoing communication when switches one network to another, results in re-issuing of request for the interrupted content retrieval at the visiting network thereby affects effective content-popularity and cache management there. Thus Wei et al. [57] proposed an optimal caching strategy that predicts the content requests by leveraging together the local geographic popularity of contents and transition probabilities (that indicate movement tendency of users). Real trace from CRAWDAD [15] was utilized to recreate interconnected networks and associated mobility while content popularity followed Zipf law (α = 0.8). It is concluded that mobility aware caching ameliorates cache hit rate as compared to mobile oblivious caching. IV. O FF -PATH C ACHING Z. Li and G. Simon [58] proposed a cooperative off-path caching strategy an economic way to handle time-shifted video streaming service keeping ISPs in mind. To achieve this authors leveraged modulo caching where each node is assigned a label that represents a positive integer value l which is smaller than a fixed integer k. A node decides to cache a chunk if the value m (equal to c modulo k) is equal to node’s label l otherwise forwards (as usual) the chunk to next hop. Results proved that proposed caching technique works better in terms of cache diversity and server load. Saucez et al. [59] to some extent shares the aim of [58] which is to minimize traffic over external links. The philosophy is to cache a content, at the maximum, at only one node to maximize cache diversity. Authors proposed CACH as a practically implementable hashing based off-path caching. For evaluation purpose homogeneous cache size, Zipf popularity distribution (α = 0.8), LRU replacement policy and real topology from Rocketfuel [12] were utilized. Results demonstrated that CACH improves hit ratio and lowers external links’ bandwidth utilization as compared to on-path caching.

In line with [59], Saha et al. [60] put forth an implicit coordinated off-path caching technique. Authors used Pathlet [61] like protocol to enable ASes to gather knowledge of its neighboring ASes along with the content range (i.e. sector of entire name space) they are willing to cache. Authors employed synthetic topology generated using BRITE [22]. ZipfMandelbrot (Z-M) distributed request pattern with α = 0.8 and q = 5 was used. Results exhibited that the proposed technique performs better in terms of off-loading server and reduction in cache replacement rate. In line with [59], Saino et al. [62] put forward off-path caching technique with different flavors of hash-routing. The paper proposed three instantiations of hash-routing based on the trajectory of retrieved contents. Authors used Icarus [30], four real topologies ( [12], [17], [19], [20]), homogeneous cache size, Zipf distributed popularity profile (0.6 ≤ α ≤ 1.1), LRU replacement policy and aggregate network cache size between 0.04% and 5% of content population. Hybrid asymmetric multicast hash-routing showed significant reduction in average load on intra-domain links by trading-off cache hit ratio marginally. In light of benefits revealed by [59] and [62], Wang et al. [63] could be viewed as yet another way to perform hash based routing dubbed as CPHR. The rationale is to segregate the total number of contents witnessed by network into different partitions (one partition per egress node). Authors used heuristic approach to map partitions to set of nodes and this mapping is shared with ingress nodes in form of Partition Assignment Table. Authors used P2P workload [36], AT&T [12] network for analysis. Results demonstrated that CPHR achieves better hit ratio as compared to default on-path caching. In view of the work done in [58], [59], [62], [63], Thar et al. [64] pointed out the issue of inconsistency in hashbased-caching when node failure or node insertion occurs. Consequently the re-mapping of hash values to the active number of nodes may result in un-balanced caches. Authors thus ventured to resolve this issue by engaging consistent hashing and forming of hash ring of virtual routers. Authors used ccnSim [25] simulator, heterogeneous caches, Zipf distribution (α ∈ [1, 1.2, 1.4, 1.6, 1.8]) and network cache size of 30% of entire content population. Results exhibited that the proposed technique performs better in terms of cache hit ratio and reduced server load. C. Li and K. Okamura [65] targeted to improve off-path caching by leveraging cluster-based hash-operated caching scheme. By permitting limited multiple replicas per content (equal to the number of clusters within an AS), internal link bandwidth usage (which is expected to be inflated due to offpath caching) is counter-balanced. Clustering therefore acts as knob to fine tune the bandwidth utilization of external and internal links. Results showed internal link bandwidth utilization is reduced for cluster based caching. H. Salah and T. Strufe [66] proposed CoMon as centrally coordinated off-path caching. A central controller ranks the content based on their popularity and prepares a contentplacement map in such a way that extremely popular contents

International Conference on Computing, Communication and Automation (ICCCA 2016) are placed at the highly central nodes and unpopular contents at the edges. The results depicted CoMon leaded to reduction in server load and hop count. V. H YBRID C ACHING Rosensweig et al. [67] developed cache-aware routing mechanism which is tailored to tap, in-general, the on-path caches and conditionally the off-path caches so as to improve response time (latency). Nodes maintain 4-tuple information referred as Breadcrum for every traversing-content. Based on that information intermediate nodes might divert the request off the path in search of the closer cached replica of the requested content. Requests followed exponentially inter arrival times and Zipf popularity distribution. Network was assumed to be congestion free. The results showed that the proposed implicit cooperative technique reduces server load and latency. Lee et al. [68] advocated coexistence of ICN with legacy TCP/IP networking by coupling both on-path and off-path caching. Such coexistence could facilitate gradual and graceful up-gradation of prevailing networking architecture to ICN. To efficiently exchange cache state summaries authors proposed use of bloom filter. Authors used topology generated by GTITM [23], Zipf popularity distribution (α = 1) and 10−2 as ratio of cache size to content population is. The proposed caching SCAN outperforms IP-routing with caching in terms of average hop count, server load and link stress. To harvest maximum benefit of collaborative caching, Guo et al. [69] coupled forwarding decisions with caching details and performed distributed estimation of content-popularity. In addition, to cope with inconsistency in popularity estimation, authors put forward self-adaptive dual-segment cache division design. Similar approach but with centralized controlling was later proposed by CoMon [66]. Authors used Abilene [14] topology, Zipf popularity distribution and two servers (in Europe and in Asia with external link latency of 120 and 200 msec respectively). The results showed that Collaborative Forwarding and Caching (CFC) works better than implicit caching of CCN, in terms of link cost and cache miss ratio. Li et al. [70] used one partition of node’s cache to hold top most popular contents independently (without communicating with other nodes) and other partition of cache to accommodate contents suggested by coordination of nodes. Authors used Zipf distribution (α ranging between 0 to 1) and real topologies (Abilene [14], CERNET [16], GEANT [17] and US-A). Results showed that as α increases higher percentage of cache size should be reserved for coordinated caching. Wang et al. [71] proposed cooperative one-hop off-path caching. A node receiving a request for a content not present in its cache, scans its one hop neighbors for the availability of the content in their caches. A bit similar approach [65] was later developed by allowing clustering within network. Authors used OmNet++ [29] simulator, ProWGen [35] to generate two traffic traces following Zipf distribution with α ∈ [0.70, 0.76, 0.90, 0.96] and shortest path routing. Both real topologies from Rocketfuel [12] and synthetic top-down topologies generated by BRITE [22] were exploited. Results

exhibited that the proposed Intra-AS cooperative scheme outperforms the default on-path caching in terms of cache hit rate and bandwidth utilization of gateway links. With the aim to transform ICN to a profitable paradigm for ISPs, Araldo et al. [72] introduced yet another dimension of cost awareness. Provided that an ISP possesses multiple external links with different cost-of-retrieval, their idea is to perform priority caching of selected contents that are being retrieved via comparatively expensive external links. Interestingly [72] exposes the two orthogonal areas of optimization in context to caching: (i) cache hit ratio optimization [41], [50], [54], [59] which aims primarily to enhance end-user experience and provides best-effort incentives to ISPs and (ii) cost aware optimization that treats service providers as first-class customers by paying equal attention to both cost-of-retrieval governed by external link used and popularity of content. Authors used Zipf popularity distribution (α ∈ [0.8, 1, 1.2]) and real topologies (from [12], [14], [17], [18]). The ratio of cache size per node to content population was varied from 0.001% to 1%. The results showed that better saving is obtained by cost aware caching. VI. C ONCLUSION In spite of several efforts made in diversified directions it seems networking community has still not reached to a common consensus regarding in-network caching. Nevertheless, work done so far interestingly reveals several facets of innetwork caching; few obvious and other unanticipated. There are couple of simulators from [25] to [31] available by now for evaluation of in-network caching in ICN. Topologies that have been frequently exploited are listed from [9] to [21] as well as tools available to generate desired topology are [22], [23], [24]. Variety of workloads (request traces) that have been used are [32], [33], [36] and the tool that has been used to generate synthetic workload is [35]. In nutshell, in-network caching needs to be deployed in a way that it should transform ICN paradigm to a profitable venture for network providers, developers and manufacturers as well as for the end users. R EFERENCES [1] V. Jacobson et al., “Networking named content,” in CONEXT’09, 2009, pp. 1–12. [2] L. Zhang et al., “Named data networking (ndn) project,” PARC, Tech. Report NDN-0001, Tech. Rep., 2010. [3] “The fp7 4ward project.” [Online]. Available: http://www.4wardproject.eu/ [4] B. Ahlgren et al., “Second netinf architecture description, 4ward eu fp7 project, deliverable d-6.2 v 2.0,” Tech. Rep., 2010. [Online]. Available: http://www.4ward-project.eu [5] N. Fotiou, D. Trossen, and G. Polyzos, “Illustrating a publish-subscribe internet architecture,” Telecommunication Systems, vol. 51, pp. 233–245, 2012. [6] N. Fotiou, P. Nikander, D. Trossen, and G. Polyzos, “Developing information networking further: From psirp to pursuit,” Broadband Communications, Networks, and Systems, pp. 1–13, 2010. [7] T. Koponen, et al., “A data-oriented (and beyond) network architecture,” SIGCOMM Comput. Commun. Rev., vol. 37, pp. 181–192, 2007. [8] “Fp7 sail project.” [Online]. Available: http://www.sail-project.eu/ [9] J. Wu, Y. Cui, X. Li, and C. Metz, “The transition to ipv6, part 1: 4over6 for the china education and research network,” IEEE Internet Computing, vol. 10, no. 3, pp. 80–85, 2006.

International Conference on Computing, Communication and Automation (ICCCA 2016) [10] “The internet topology data kit - 2011-04.” [Online]. Available: http://www.caida.org/data/active/internet-topology-data-kit [11] A. L. Barabasi and R. Albert, “Emergence of scaling in random networks,” science, vol. 286, no. 5439, pp. 509–512, 1999. [12] N. Spring, R. Mahajan, and D. Wetherall, “Measuring isp topologies with rocketfuel,” in ACM SIGCOMM Computer Communication Review, vol. 32, no. 4, 2002, pp. 133–145. [13] D. J. Watts and S. H. Strogatz, “Collective dynamics of ‘small-world ’networks,” Nature, vol. 393, no. 6684, pp. 409–410, 1998. [14] Abilene, “Internet2 ip igp metrics.” [Online]. Available: http://noc.net.internet2.edu/i2network/mapsdocumentation/maps.html [15] I. Rhee et al., “Crawdad data set ncsu/mobilitymodels (v. 2009-07-23),” July, 2009. [Online]. Available: http://crawdad.cs.dartmouth.edu/ncsu/mobilitymodels [16] “The cernet topology.” [Online]. Available: http://www.edu.cn/20060111/3170194.shtml [17] “The geant topology.” [Online]. Available: http://www.geant.net [18] Tiger, “Topology integrated geographic encoding and referencing system.” [Online]. Available: http://www.census.gov/geo/www/tiger [19] GARR, “Italian academic network - consortium garr.” [Online]. Available: www.garr.it [20] WIDE, “Japanese academic network.” [Online]. Available: http://two.wide.ad.jp [21] “Planetlab.” [Online]. Available: http://www.planetlab.org [22] “Brite.” [Online]. Available: http://www.cs.bu.edu/brite/ [23] E. W. Zegura, “Gt-itm: Georgia tech internetwork topology models (software), georgia tech,” 1996. [Online]. Available: http://www.cc.gatech.edu/fac/Ellen.Zegura/gt-itm/gt-itm/tar.gz [24] J. Winick and S. Jamin, “Inet-3.0: Internet topology generator,” Technical Report CSE-TR-456-02, University of Michigan, Tech. Rep., 2002. [25] “ccnsim.” [Online]. Available: http://www.enst.fr/ drossi/ccnSim [26] “ccnx project.” [Online]. Available: http://www.ccnx.org [27] “Opnet.” [Online]. Available: http://www.opnet.com/ [28] “Ccnpl.” [Online]. Available: https://code.google.com/p/ccnpl-sim/ [29] “Omnet++.” [Online]. Available: http://www.omnetpp.org [30] “Icarus.” [Online]. Available: http://www.ee.ucl.ac.uk/ lsaino/software/icarus [31] A. Quereilhac, A. Kalla, T. Turletti, and W. Dabbous, “Easy ccnx experimentation on planetlab,” 2012. [32] I. Cantador, P. Brusilovsky, and T. Ku?ik, “Second workshop on information heterogeneity and fusion in recommender systems (hetrec2011).” in RecSys, 2011, pp. 387–388. [33] J. Leskovec and J. McAuley, “Learning to discover social circles in ego networks,” in Advances in neural information processing systems, pp. 539–547, 2012. [34] K. Wang et al., “Content-centric networking: Effect of content caching on mitigating dos attack,” International Journal of Computer Science, vol. 9, no. 3, pp. 43–52, 2012. [35] M. Busari and C. Williamson, “Prowgen: a synthetic workload generation tool for simulation evaluation of web proxy caches,” Computer Networks, vol. 38, no. 6, pp. 779–794, 2002. [36] M. Hefeeda and B. Noorizadeh, “On the benefits of cooperative proxy caching for peer-to-peer traffic,” IEEE Transactions on Parallel and Distributed Systems, vol. 21, no. 7, pp. 998–1010, 2010. [37] G. Carofiglio, V. Gehlen, and D. Perino, “Experimental evaluation of memory management in content-centric networking,” in Communications (ICC), IEEE International Conference on, 2011, pp. 1–6. [38] D. Rossi and G. Rossini, “Caching performance of content centric networks under multi-path routing,” Telecom ParisTech, Tech. Rep., 2011. [39] M. Gallo et al., “Performance evaluation of the random replacement policy for networks of caches,” in ACM SIGMETRICS Performance Evaluation Review, vol. 40, no. 1, pp. 395–396, 2012. [40] D. Rossi and G. Rossini, “On sizing ccn content stores by exploiting topological information,” in INFOCOM Workshops, 2012, pp. 280–285. [41] W. K. Chai, H. Diliang, I. Psaras, and G. Pavlou, “Cache “less for more” in information-centric networks,” in NETWORKING, pp. 27–40, 2012. [42] Y. Wang et al., “Optimal cache allocation for content-centric networking,” in Network Protocols (ICNP), 21st IEEE International Conference on, 2013, pp. 1–10. [43] Z. Ming, M. Xu, and D. Wang, “Age-based cooperative caching in information-centric networks,” in INFOCOM Workshops, 2012, pp. 268– 273.

[44] K. Cho et al., “Wave: Popularity-based and collaborative in-network caching for content-oriented networks,” in INFOCOM Workshops, 2012, pp. 316–321. [45] S. J. Kang, S. W. Lee, and Y. B. Ko, “A recent popularity based dynamic cache management for content centric networking,” in Ubiquitous and Future Networks (ICUFN), Fourth International Conference on, 2012, pp. 219–224. [46] G. Tyson et al., “A trace-driven analysis of caching in content-centric networks,” in Computer Communications and Networks (ICCCN), 21st International Conference on, 2012, pp. 1–7. [47] S. Wang, J. Bi, and J. Wu, “On performance of cache policy in information-centric networking,” in Computer Communications and Networks (ICCCN), 2012 21st International Conference on, pp. 1–7. [48] I. Psaras, W. K. Chai, and G. Pavlou, “Probabilistic in-network caching for information-centric networks,” in 2nd ACM SIGCOMM ICN Workshop, 2012, pp. 55–60. [49] J. M. Wang and B. Bensaou, “Progressive caching in ccn,” in IEEE GLOBECOM, 2012, pp. 2727–2732. [50] S. K. Fayazbakhsh et al., “Less pain, most of the gain: incrementally deployable icn,” in ACM SIGCOMM, 2013, pp. 147–158. [51] A. Dabirmoghaddam, M. M. Barijough, and J. J. Garcia-Luna-Aceves, “Understanding optimal caching and opportunistic caching at ”the edge” of information-centric networks,” in ACM ICN, 2014, pp. 47–56. [52] J. Li et al., “Popularity-driven coordinated caching in named data networking,” in Proc. of the eighth ACM/IEEE symposium on Architectures for networking and communications systems, 2012, pp. 15–26. [53] C. Bernardini, T. Silverston, and O. Festor, “Socially-aware caching strategy for content centric networking,” in Networking Conference, IFIP, 2014, pp. 1–9. [54] M. Badov et al., “Congestion-aware caching and search in informationcentric networks,” in Proceedings of the 1st international conference on Information-centric networking, 2014, pp. 37–46. [55] D. Nguyen, K. Sugiyama, and A. Tagami, “Congestion price for cache management in information centric networking,” in INFOCOM Workshops, 2015, pp. 287–292. [56] G. Carofiglio, L. Mekinda, and L. Muscariello, “Lac: Introducing latency-aware caching in information-centric networks-technical report,” arXiv preprint arXiv:1506.06642, Tech. Rep., 2015. [57] T. Wei, L. Chang, B. Yu, and J. Pan, “Mpcs: A mobility/popularity-based caching strategy for information-centric networks,” in IEEE GLOBECOM, 2014, pp. 4629–4634. [58] Z. Li and G. Simon, “Time-shifted tv in content centric networks: the case for cooperative in-network caching,” in Communications (ICC), 2011 IEEE International Conference on, 2011, pp. 1–6. [59] D. Saucez, A. Kalla, C. Barakat, and T. Turletti, “Minimizing bandwidth on peering links with detection in named data networking,” INRIA Sophia Antipolis Mediterranee, Tech. Rep., April 2012. [Online]. Available: http://hal.inria.fr/hal-00684453 [60] S. Saha, A. Lukyanenko, and A. Yla-Jaaski, “Cooperative caching through routing control in information-centric networks,” in INFOCOM, 2013, pp. 100–104. [61] P. Godfrey, I. Ganichev, S. Shenker, and I. Stoica, “Pathlet routing,” SIGCOMM Comput. Commun. Rev., vol. 39, no. 4, pp. 111–122, 2009. [62] L. Saino, I. Psaras, and G. Pavlou, “Hash-routing schemes for information centric networking,” in Proc. of the 3rd ACM SIGCOMM Workshop on Information-centric networking, 2013, pp. 27–32. [63] S. Wang, J. Bi, and J. Wu, “Collaborative caching based on hash-routing for information-centric networking,” in ACM SIGCOMM Computer Communication Review, vol. 43, no. 4, 2013, pp. 535–536. [64] K. Thar, S. Ullah, and C. S. Hong, “Consistent hashing based cooperative caching and forwarding in content centric network,” in Network Operations and Management Symposium (APNOMS), 2014, pp. 1–4. [65] C. Li and K. Okamura, “Cluster-based in-networking caching for content-centric networking,” International Journal of Computer Science and Network Security (IJCSNS), vol. 14, no. 11, 2014. [66] H. Salah and T. Strufe, “Comon: An architecture for coordinated caching and cache-aware routing in ccn,” in Consumer Communications and Networking Conference (CCNC), 2015 12th Annual IEEE. IEEE, 2015, pp. 663–670. [67] E. J. Rosensweig and J. Kurose, “Breadcrumbs: Efficient, best-effort content location in cache networks,” in INFOCOM. IEEE, 2009, pp. 2631–2635.

International Conference on Computing, Communication and Automation (ICCCA 2016) [68] M. Lee, K. Cho, K. Park, T. T. Kwon, and Y. Choi, “Scan: Scalable content routing for content-aware networking,” in Communications (ICC), 2011 IEEE International Conference on, 2011, pp. 1–5. [69] S. Guo, H. Xie, and G. Shi, “Collaborative forwarding and caching in content centric networks,” in NETWORKING, 2012, pp. 41–55. [70] Y. Li, H. Xie, Y. Wen, and Z. L. Zhang, “Coordinating in-network caching in content-centric networks: Model and analysis,” in Distributed Computing Systems (ICDCS), 2013 IEEE 33rd International Conference on, 2013, pp. 62–72. [71] J. M. Wang, J. Zhang, and B. Bensaou, “Intra-as cooperative caching for content-centric networks,” in Proceedings of the 3rd ACM SIGCOMM Workshop on Information-centric networking, 2013, pp. 61–66. [72] A. Araldo, D. Rossi, and F. Martignon, “Cost-aware caching: Caching more (costly items) for less (isps operational expenditures),” IEEE Transactions on Parallel and Distributed Systems, 2015.