Living in a PIT-less World: A Case Against Stateful Forwarding in Content-Centric Networking
Information-Centric Networking (ICN) is a recent paradigm that claims to mitigate some limitations of the current IP-based Internet architecture. The centerpiece of ICN is named and addressable content, rather than hosts or interfaces. Content-Centric Networking (CCN) is a prominent ICN instance that shares the fundamental architectural design with its equally popular academic sibling Named-Data Networking (NDN). CCN eschews source addresses and creates one-time virtual circuits for every content request (called an interest). As an interest is forwarded it creates state in intervening routers and the requested content back is delivered over the reverse path using that state.
Although a stateful forwarding plane might be beneficial in terms of efficiency, and resilience to certain types of attacks, this has not been decisively proven via realistic experiments. Since keeping per-interest state complicates router operations and makes the infrastructure susceptible to router state exhaustion attacks (e.g., there is currently no effective defense against interest flooding attacks), the value of the stateful forwarding plane in CCN should be re-examined.
In this paper, we explore supposed benefits and various problems of the stateful forwarding plane. We then argue that its benefits are uncertain at best and it should not be a mandatory CCN feature. To this end, we propose a new stateless architecture for CCN that provides nearly all functionality of the stateful design without its headaches. We analyze performance and resource requirements of the proposed architecture, via experiments.
Information-Centric Networking (ICN)  is a networking model that emerged as an alternative to the host-based communication approach of the current IP-based Internet architecture. Content-Centric Networking (CCN) [2, 3] is one industry-driven instance of this model. (It is closely related to Named-Data Networking (NDN), which can be viewed as CCN’s academic dual.) While IP traffic consists of packets sent between communicating end-points, CCN traffic is comprised of explicit requests for, and responses to, named content objects. These requests, called interest messages, refer to the desired content by name. An interest is forwarded by routers (using the name) towards a content producer until satisfied by the latter or by a cached copy in some router. The corresponding response, called content, is forwarded along the reverse path. To reduce end-to-end latency and congestion, CCN routers may opportunistically cache content to satisfy future interests.
In CCN, neither interest nor content messages carry source addresses. In order to correctly deliver content to consumers, routers maintain per-interest state as an entry in a so-called Pending Interest Table (PIT). This state information maps interest names to interfaces on which they arrived. This enables a router which receives a content easily identify the forwarding interface(s) using the corresponding PIT entry. Once a content is forwarded downstream, the corresponding PIT entry is flushed.
Another purpose of the PIT is to support interest collapsing – a feature useful for handling multitudes of nearly simultaneous interests for the same content. Whenever a router receives an interest for which it has a matching PIT entry, the arrival interface of the new interest is added to the existing entry and the interest is not forwarded further. This prevents duplicate interests from being sent upstream, thus lowering overall congestion. However, as we show later in the paper, interest collapsing rarely occurs in practice.
Furthermore, stateful forwarding enabled by PITs is supposed to provide flow balance via path symmetry between interest and content messages. Consequently, information from the PIT (e.g., interest to content Round-Trip Time (RTT)) can be used to develop better congestion control and traffic shaping mechanisms; see [4, 5, 6, 7]. However, using a PIT for flow balance and in-network congestion control is quite problematic in practice. In fact, flow balance is a false claim in the current CCN design due to the (potentially huge) disparity in sizes between interest and content messages. Likewise, there is ample evidence that congestion control and transport protocols are best deployed at the receiver [8, 9, 10], due to flow imbalance and dynamic routing in CCN. Another claimed advantage of stateful forwarding is that it can aid routers when responding to network problems since they can make autonomous and intelligent forwarding decisions for interests. In practice, however, individual routers rarely have sufficient autonomy to make such decisions.
From the perspective of infrastructure security, the PIT effectively prevents reflection attacks since content is always forwarded according to PIT entries . However, such attacks can be mitigated by mechanisms that do not require any forwarding state. Moreover, this state is costly to maintain. Various attempts to improve the efficiency of PIT-based forwarding have been studied in the context of CCN and NDN [12, 13, 14]. However, they do not address the fundamental design issue that the PIT size grows linearly with the number of distinct interests received by a router. This means that a PIT is a resource that can be easily abused. In fact, malicious exhaustion of PIT space in the form of Interest Flooding (IF) attacks  remains an important open problem. In such attacks, adversaries can flood routers with nonsensical (i.e., unsatisfiable) or slow-to-satisfy (i.e., requiring dynamic content generation) interests in order to maximize the occupied PIT space. Once a router reaches its maximum PIT capacity it either: (1) drops new incoming interests, or (2) removes existing entries to free resources for new incoming interests. Unfortunately, these two options result in effective DoS for legitimate future or current interests, respectively.
Given that many of the claimed benefits are dubious and considering associated the infrastructure security problems, it becoms hard to justify the need for PITs in CCN. Therefore, in this paper, we comprehensively assess (in Section II) the stateful forwarding plane of CCN with respect to each claimed benefit. We show that such benefits are: (1) either unrealistic or infeasible in practice, (2) can be achieved by means other than stateful forwarding, or (3) so marginal that their worth simply does not justify the overhead. We then present, in Sections III and IV, a new stateless architecture for CCN based on Routable Backward Names (RBNs). This new design can co-exist with the current CCN architecture (with PITs) or replace it entirely. Experimental results in Section V indicate that the new design still retains the essence and performance characteristics of CCN while successfully avoiding pitfalls of stateful forwarding. We conclude with a discussion of related work and a summary in Sections VI and VII, respectively.
Ii Assessing the PIT
The PIT is a fundamental and mandatory feature of the CCN forwarding plane. It is a tabular data structure that maps interest names and other metadata to a set of information, such as arrival interfaces and lifetime values. Arrival interfaces are used to identify downstream interfaces on which content responses should be forwarded. The shape and size of this table is directly dependent on the traffic that is processed by a forwarder.  studied dynamics of the PIT and showed that the number of entries can range from less than for edge routers with a small number of per-namespace flows to over in the core.  designed a PIT implementation that requires only 37MiB to 245MiB to forward traffic at 100Gbps, which can scale to fit the needs of realistic traffic, according to .
However, in this paper, we do not question the implementaton of the PIT. Instead, we question its entire existence. Below, we argue that aside from being unecessary to support CCN-like communication, the PIT’s presence raises more (serious) problems than it solves. We support our argument by systematically analyzing the following alleged PIT benefits:
Content object forwarding
Flow and congestion control
We then show that all these benefits are either false, unnecessary, or very meager at best.
Ii-a Content Object Forwarding
A key tenet of CCN is that content is never sent to a consumer who did not previously issue an interest for the (name of that) content. Since interests contain no source addresses, PITs are needed “[t]o forward Content Objects from producers to consumers along the interest reverse path by leaving per-hop state in each router…” .
We disagree with this statement for two reasons. First, network path symmetry is not guaranteed and should not be assumed. Wolfgang et al.  demonstrated that route symmetry between the same flow on the Internet is lower in the core than at the edges. Several tier-1 and tier-2 networks were studied and it was shown that, due to “hot-potato-routing,” flow asymmetry exceeds in the core. Thus, symmetric path routing in the core appears to direct contradict today’s practices that promote path asymmetry. Attempting to enforce symmetric data traversal appears to be a challenge from an economic perspective.
Second, pull-based communication with symmetric paths is not well-suited for all applications. While appropriate for scalable content distribution applications111Which some believe to be already well-served by today’s CDNs., it is substantially different from modern TCP/IP applications and protocols which rely on point-to-point bi-directional streams between endpoints. For instance, the WebSocket  protocol uses full-duplex TCP streams for clients and servers that engage in real-time, bidirectional communication. It is used by many popular interactive applications, such as multimedia chat and multiplayer video games. Two-way communication is not limited to Web protocols. Voice applications such as Skype  and peer-to-peer systems such as BitTorrent  rely on two endpoints which both produce and consume data, as part of the application.
Given the relative infancy of CCN and abundance of real-world applications that currently do not fit CCN’s mold, it is difficult to argue that the pull communication model can satisfy all applications’ needs. For example, even some existing CCN applications exploit interest messages to carry information from consumers to producers . Other applications rely on consumers and producers to send interests to each other. NDN-RTC, a recently developed NDN video teleconference application, is one specific example that supports such bidirectional communication between peers . (We use NDN and CCN interchangeably here since both are equivalent in this context.) Another emerging application design pattern is data transport via set synchronization, notably, the NDN ChronoSync protocol . It enables data synchronization among a set of users. Each ChronoSync user acts as both a producer and consumer. Consumers (members) issue long-standing interests to a group (common namespace) about specific data to be synchronized, that are routed to all members. When target data is changed by someone, this member satisfies previous interest(s) with a fingerprint of the data in a content object. Each member is then responsible for requesting updated content to synchronize with the others.
Based on the trends of current TCP/IP applications and proposed design strategies for CCN-based protocols and applications, it seems clear that bidirectional communication is here to stay. For it to work, router FIBs need to contain prefixes for all end-points – not just producers. Therefore, all communicating parties need to obtain and use a routable prefix, which effectively serves as an address. As a consequence, forwarding information stored in a PIT becomes redundant and unnecessary.
Ii-B Uitility of Interest Collapsing
Recall that collapsing applies only to interests arriving at routers during a very small time-window : between the time of arrival of the original interest (referring to the same name) that triggered creation of a new PIT entry, and the time of arrival of the corresponding content. Due to increasing network data rates and lower end-to-end delays facilitated by in-network caching, the value of is expected to be miniscule, e.g., on order of tens of milliseconds. Therefore, we believe that the effect of interest collapsing would not play any significant role in the real life performance of CCN.
To support this claim, we model the probability of interest collapsing occurring in the first-hop router . The reason for choosing the first router is that the benefits of interest collapsing are felt the most closest to the consumer(s). This is because collapsing two similar interests at the consumer-facing router reduces bandwidth usage more than when the same occurs closer to the producer. We assume that content popularity follows a Zipf distribution with classes and average number of segments .222Large content is typically split into smaller segments. Let each class arrival rate at be modeled as a Poisson process. The event of interest collapsing at for content class is denoted as . The probability of this event is :
Assuming in-network routing is only enabled at edge routers , the interest collapsing probability at consumer-facing router is:
for links between and producer , the speed of light, the length of link , constant that depends on the characteristics of the link’s physical material, and the cache hit probability of a class content at .
We focus on modeling interest collapsing for individual content objects. Thus, we set content size segment. Therefore, Equation 1 can be re-written as follows.
However, taking into consideration content caching at , the previous equation can be further re-written as:
where is the cache miss probability. In other words, if requested content is cached, satisfies corresponding interests without creating PIT entries and interest collapsing does not occur.
We now redefine as a function of propagation delays on each link on the path: . is the propagation delay of the link between and , where and . Moreover, represents cache hit probability at and, if an interests generates a cache hit at , it is not propagated further.
where represents the propagation speed of link , and is the index of router where a cache hit first occurs.
However, assuming that in-network caching only happens at the edges and that is negligible relative to (we ignore the effect of caching at ), cache hit probability in all routers between and (not including ) is zero, and:
This concludes the proof. We analyze Theorem 1 in the following setup. For simplicity’s sake, we use Equation 3 for content arrival rates and propagation delays between and . Since content popularity follows a Zipf distribution, arrival rate for class is half of that for class , i.e., . To illustrate the highest possible interest collapsing probability, we assume that requested content (even if popular) is not cached at . Figure 1(a) shows the collapsing probability of four content classes . The graph only considers propagation delay up to milliseconds because, as shown in , the virtual RTT (VRTT)333RTT taking into consideration existence of caches. for content class is around milliseconds. We note that for the most popular content . However, in a more realistic setup where ’s cache is taken into consideration, the highest interest collapsing probability is for content class ; see Figure 1(b). Based on such low probabilities, we conclude that interest collapsing is not crucial for a content distribution network such as CCN.
Ii-C Flow and Congestion Control
Yi et al.  presented the first thorough argument in support of a stateful forwarding plane in the context of NDN. Due to their near-identical features, the same applies to CCN. The PIT can be used to record RTTs for interest and content exchanges, which, in turn, is useful for making dynamic forwarding decisions. For instance, if the RTT for a given namespace on a particular link becomes too high, that link might be congested and alternatives should be explored. This type of in-network congestion and flow control has been studied further in [4, 26, 10]. For example,  propose a joint hop-by-hop (i.e., in-network) and receiver-based control protocol that relies on PIT-based RTT measurements for flows.
However, according to , flow differentiation is a difficult challenge. Thus, one approach to “interest shaping” is by controlling the flow of data on upstream and downstream links independently of flows. This does not require any information from the PIT. Instead, it relies on knowledge of average interest and content size, link bandwidths, and interest arrival rates (or demand). Similar to , it also relies on receiver-driven flow control via an Additive-Increase-Multiplicative-Decrease window.  is another example of a receiver-driven flow control protocol for CCN. Given these results, current trends in the ICN research community favor pushing stateful control protocols to receivers, rather than to network nodes.
Ii-D Infrastructure Security
Denial of Service (DoS) attacks are a major threat to any network infrastructure. DoS attacks in today’s Internet include: bandwidth depletion, DNS cache poisoning, black-holing and prefix hijacking, as well as reflection attacks. Gasti et al.  show how CCN (in the context of NDN) prevents these types of attacks. Out of all attack types considered, the PIT is needed only to prevent reflection attacks . Since content is forwarded based on PIT entries such attacks are impossible in CCN. However, forwarding content via the PIT is not the only way to prevent reflection attacks. If packets have a source address, the ingress filtering technique in  – whereby ISPs filter packets based on source addresses – would work equally well.
Despite its resilience to reflection attacks, CCN is susceptible to another major attack type known as Interest Flooding (IF) . In one IF attack flavor, a malicious consumer (or a distributed botnet) issues nonsensical interests444For example, an interest with a name reflecting a valid producer’s prefix, with a random number as its last component. so as to overwhelm targeted routers and saturate their PITs. According to , the PIT size can exceed as upstream paths become congested. The problem worsens if a malicious consumer and producer cooperate to target a specific router. Although several attempts to detect, mitigate, and prevent them have been made [31, 32, 33, 34, 35, 36, 37]555For details, see Section VI below. each of them is effective against only a very naïve or weak attacker. Thus, IF attacks remain a big open problem with no comprehensive solution in sight.
Iii Stateless CCN using Backwards Routable Names
Based on our earlier discussion, the price of a PIT comes at the price of serious infrastructure security problems that have not been addressed. To this end, we introduce a modified stateless CCN architecture, called stateless CCN.
The main idea behind our stateless CCN design is simple: an interest now includes a new field called Backwards Routable Name (BRN), a routable prefix, similar to an IP source address. A BRN indicates where the corresponding content should be delivered, akin to an IP destination address. The corresponding content carries the BRN as its routable name towards the consumer. Thus, with properly configured FIB entries, content is correctly delivered to the requesting consumer.666This requires consumers to publicly advertise their BRN prefixes and participate in routing. This modification to the CCN architecture is clearly inspired by IP – all packets (interest and content) are forwarded based on addresses they carry and not on network state. However, as we show below, this does not violate CCN’s core value of named data being moved through, and stored in, the network.
To illustrate BRN-based forwarding, consider a scenario where a consumer with name lci:/edu/uci/ics/bob () requests content from a producer with the name lci:/bbc/news/today ().777Names are encoded using the Labeled Content Identifier (LCI) schema . LCI names are the concatenation of individual name components, separated by the ‘/’ character, in a typical URI-like format. Let be an interest with the Routable Name and Supporting Name . Also, let be the corresponding content object that matches where , and . In this example, assume that is not cached anywhere.
advertises its name and the routing protocol propagates this information accordingly.
The network forwards to according to router FIB entries.
Once receives it replies with .
Similarly to Step 3, the network forwards back to using the same interest forwarding strategy.
Several modifications need to be made to the existing CCN architecture and protocol to enable this communication. At a minimum, interest and content object messages should carry two names: one of the requested content and the other of the requesting consumer. These two names corresponds to source and destination IP addresses in today’s Internet.
We suggest modifying both interest and content headers to include a new field called SupportingName (SN). This field contains the BRN of the interest-issuing consumer. In the above example, interest header would contain lci:/cnn/news/today and lci:/edu/uci/ics/bob as and , respectively. The replied content header would contain the same and values. Note that content object signatures can be generated in advance by omitting the content’s field since this is only used for routing purposes. The resulting packet formats are shown in Figure 2 in ABNF form.
Currently, interest and content messages are very similar in CCN. Both contain a Name, Payload, and optional Validation fields ; they only differ in the top-level type. Our stateless variant still requires this distinction since interests and content objects are processed differently. For example, a router first attempts to satisfy an interest from its cache, while content is (optionally) cached prior forwarding.
We stress that a content might not follow the reverse path of the proceeding interest due to routing table configurations. In fact, we anticipate that consumers might structure BRNs to control the degree of path asymmetry between interest and content messages.
Modified interst and content formats coupled with PIT removal simplify router’s fast-path processing. Algorithms 1 and 2 show how a router would process interest and content messages. CS-Lookup represents a CS lookup operation based on (content name). For clarity’s sake, we omit content verification details in all algorithms.
Iv Architecture Assessment
Despite significant research progress over the past 5 years, the PIT no longer seems to be a practical solution for content object forwarding in CCN. As discussed earlier, router PITs are prone to DoS (specifically, IF) attacks. They also store information already available from FIBs (consumer routable prefixes) and enforce unnatural path symmetry in an increasingly asymmetric Internet. Moreover, flow and congestion control algorithms are being pushed towards receivers instead of in the network based on state maintained in PITs. Our simple stateless CCN variant mitigates these problems by specifying the use of source and destination prefixes. To support our claims, we compare the stateful and stateless CCN architectures with respect to aforementioned features. We then discuss both advantages and disadvantages of stateless CCN.
Iv-a Revisiting PIT Benefits
Reverse-Path Routing. Our stateless CCN scheme requires FIBs to be updated to accommodate RBN prefixes advertised by consumers. It might seem that this would lead to a tremendous increase in FIB size. However, recall that CCN interest (and now, content) forwarding is based on LPM. In stateless CCN, consumers announce their RBNs only to their first-hop routers (e.g., an access point), which, in turn, combines all its consumers’ RBNs and announces an aggregate prefix to neighboring routers, similar to the Border Gateway Protocol (BGP) route-aggregation feature .
Also, path asymmetry between interest and content messages in stateless CCN is more compliant with networking and routing practices of today’s Internet. As argued in Section II-A, ISPs are likely to adopt an architecture that agrees with their present business model.
Forwarding Overhead. Stateful CCN dictates that, when processing an interest, a router should, in the worst case: (1) attempt to satisfy the interest from its cache, (2) create or modify a PIT entry for the interest, and (3) perform a FIB lookup. Meanwhile, stateless CCN eliminates (2), which reduces router operations to cache and FIB lookups.
Removing the PIT also simplifies and improves content forwarding logic. Instead of indexing the PIT to obtain downstream interfaces, a router performs a FIB lookup for each content, just as it does in processing an interest. We claim that a simple FIB lookup is more efficient than using the PIT for forwarding content. After a PIT lookup, a PIT entry is flushed once a content object is forwarded. This ”flush” incurs an additional write operation in the routers’ fast path. Stateless CCN replaces this with a simple LPM-based FIB lookup. Note that LPM algorithms have been intensively studied, constructed, and fine-tuned to cope with multi-gigabit, and even terabit, IP packet processing [41, 42, 43].
Flow and Congestion Control. The current receiver driven flow and congestion control algorithms are unaffected in our stateless variant. The only difference is that now routers are unable to compute the RTT for a given interest-content exchange. Given that recent algorithms do not rely on these calculations in the network anyway, this is a tolerable loss.
Iv-B Content Caching
As mentioned earlier, using RBNs for content routing does not preserve path symmetry. In fact, it encourages path asymmetry. Consequently, content might be cached along a different path than the interest originally traversed. It might seem that adjacent (or nearby) consumers for the same content would therefore not benefit from in-network caching. We argue that this is not so.
Firstly, CCN content includes a producer-specified cache hint that suggests how long routers ought to cache this content. Routers are expected to honor this hint when managing the data in their caches. Secondly, recall that routers can unilaterally decide whether to enable caching for none, some, or all content. Thirdly, cache eviction strategies might reduce cache entry lifetime to much less than the suggested value. For instance, core routers would most likely not cache content given their high processing rates. Meanwhile, consumer-facing routers would handle much less traffic and are thus more likely to cache content for the required amount of time. In fact, caching has been shown to be most cost effective at the edges , e.g., at the tier-3 ISP level. Since nearby consumers share the same edge router, they will all benefit from caching popular content in that router. This observation is supported by the results obtained in , wherein it is shown that path symmetry is highest at the edges of the network.
Figure 3 shows an example of caching in stateless CCN. The topology has 4 autonomous systems (AS-s). AS1 and AS4 are stubs representing tier-3 ISPs, while AS2 and AS3 are transits representing tier-1 ISPs.888We ignore tier-2 ISPs for simplicity. Interests issued by are forwarded towards along the dotted (red) path, and content is forwarded back to along the dashed (blue) path. Assuming that caching only occurs near the edges, content sent from to gets cached in AS4. Consequently, interests for the same content issued by other consumers in AS4 would be satisfied from cache(s) of AS4.
Iv-C Infrastructure Security
We now discuss both beneficial and problematic infrastructure security issues in stateless CCN, such as (D)DoS attacks.
Interest Flooding. Stateless CCN mitigates this attack by eliminating its root cause – the PIT. Without per-request state in routers, this attack vector is removed. This represents the major advantage of stateless CCN.
Reflection Attacks. Interest and content path symmetry in CCN prevents reflection attacks. However, in stateless CCN, RBNs serve as a de facto source address in interest, and destination in content, messages. Thus, reflection attacks re-appear. Fortunately, the ingress filtering technique described in  can be used to mitigate them.
Cache and Content Poisoning. Content authentication in stateless CCN is identical to that in the stateful CCN architecture. It is done by producers signing content objects or using Self-Cerftifying Names (SCN) . Regardless of the method, all content must be verified by consumers. However, verification is not mandatory for routers, for several reasons; see  for more details. Lack of in-network content verification opens the door for content poisoning attacks . Moreover, due to possible path asymmetry in RBN-based content forwarding, content poisoning countermeasures that work in the current CCN architecture do not apply anymore.
The PIT enables a router to apply the so-called Interest-Key Binding (IKB) rule , whereby consumers and producers collaborate to provide routers with enough (minimal) trust information to perform content verification. This information is currently stored in the PIT. However, as mentioned above, path asymmetry renders IKB impractical. In stateless CCN, a router might receive (unsolicited) content without prior interest traversing the same path. If such content is returned on a path different from the original interest, routers cannot trust any information it carries.
Consumer Privacy. Lack of source addresses in stateful CCN facilitates a degree of consumer privacy. RBNs in stateless CCN negate this benefit. However, given highly descriptive nature of content names, end-to-end encryption might be needed to achieve better privacy, even in stateful CCN. Moreover, since many current applications customize content per consumer, both interest and content messages, if left unencrypted, might include enough information to identify the consumer. Thus, for sensitive or consumer-specific content, stateless CCN does not necessarily yield worse privacy due to the use of encryption.
Beyond naming issues, in-network caching can be abused by an adversary to compromise both consumer and producer privacy . By measuring the time required for content retrieval, an adversary can learn whether specific content was recently requested by other nearby consumers. This attack is still applicable in stateless CCN since in-network caching remains a feature. Fortunately, countermeasures proposed in [46, 47] are equally effective in both architectures.
Iv-D Deployment Issues
The main purpose of designing a stateless CCN architecture is to provide an alternative to the current stateful CCN. This does not mean that one must replace the other. In fact, they can co-exist and allow the consumers to select one or the other on-demand. Consider the following scenarios:
Stateless CCN: includes an RBN () in an interest and upstream routers forward it as necessary. Stateful routers create PIT entries and stateless routers do not. In both cases, the interest is forwarded according to the FIB using content name . Upon receipt of a content message, a stateful router uses its PIT to forward the content downstream, while a stateless router does that using the FIB and . In this case, stateful forwarders simply ignore the fields in both interests and content objects. This makes the proposed stateless CCN backwards compatible with the current CCN architecture.
Stateful CCN: issues an interest per current CCN rules. If a stateless router receives such an interest, it generates a NACK indicating that the interest cannot be forwarded further. To handle this NACK, some downstream node must provide a RBN for the interest and re-forward it as needed. This node can be an AS gateway (i.e., a router that can forward packets to and from other ASs) or, worst case, the consumer.
Any node that satisfies an interest must honor its version (stateless or stateful) when producing a response. For example, if a producer (or a caching router) receives an interest with an RBN, it must reply according to stateless CCN by keeping both and in the corresponding content.
We claim that this type of hybrid approach aligns very well with CCN’s edge-caching strategy  and current path asymmetry in the Internet’s core. Recall that stateless CCN makes it impossible for stateless routers to verify content they forward. Moreover, if a router cannot verify a content then it should not cache it. Thus, caching would only occur near the edges where PITs are located. We envision a network where all routers within an AS have PITs that are used for egress interests, i.e., those generated by consumers within the AS. Traditional verification techniques can be applied at these routers for returned content. If an interest leaves an AS, the gateway first supplies a RBN before forwarding it upstream. This interest will not induce any PIT state upstream and therefore will not result in the corresponding content object being cached outside of the AS. This is not problematic though since the results of  imply that caching near the edge (i.e., within the AS above) is most effective. Moreover, this approach allows path asymmetry outside of the AS, which aligns with the real-world routing strategies noted in .
Another side-effect of the hybrid approach is that it can be used as an IF attack recovery mechanism. If implements a PIT but does not have enough resources to create a new entry for , can respond with a NACK similar to what is described above. In this case, if is under an IF attacks and its PIT resources is exhausted, neither current nor future interests will be dropped. The disadvantage of this approach as an effective IF attack countermeasure is that (1) it is reactive, so it can only be used after the attack occurs, and (2) it incurs an additional end-to-end latency since consumers (or downstream routers) need to reissue the interests following stateless CCN guidelines.
V Experiments and Analysis
We now evaluate performance of the stateless CCN in relation to stateful CCN. Our key metric is the degree to which forwarding overhead is lowered by stateless routing. To do this, we modified the ndnSIM  simulator, a simplified NDN implementation as a NS-3  module, to support the stateless CCN architecture proposed in Section III. Specifically, we modified the NDN Forwarding Daemon (NFD)  to supporting interest forwarding based on content names and content forwarding based on RBNs, without leaving PIT states behind.
We then simulated topologies based on the Deutsches ForschungsNetz (DFN), the German Research Network [51, 52], and AT&T networks (shown in Figures 4(a) and 4(b), respectively). Each topology consists of consumers999Each consumer node in the figures consists of actual consumers., a single producer connected to one of the edge routers, and multiple routers (more than ). Each consumer generates interests per second, with a random suffix so as to avoid cache hits. This is done to force interest to traverse all the path to the producer, hence maximize the amount of processing that takes place in the forwarders in both the upstream and downstream paths. This captures the worst-case scenario.
Our results are shown in Figure 5. In both topologies, we observe approximately improvement in per packet forwarding performance of interest messages. We also observe a improvement in processing content objects. These cost savings are quite significant, especially, for core routers that might process packets at rates of Gbps and over.
Furthermore, the overall content retrieval latency is significantly improved when using stateless forwarders as compared to stateful ones. Figure 6 shows a comparison of the RTT performance for both forwarders in the DFN topology. In this experiment, consumers always request unique content in order to avoid cache hits.101010We do not take caching into consideration to eliminate any randomize effect it might have on content retrieval latency. Such randomized effects can be cause by different cache eviction policies.. Figure 6 illustrates the average RTT to fetch content objects as a function of the number of hops between consumers and the producer. We notice a content retrieval latency improvement of more than . This improvement reaches for paths consisting of hops.
To summarize, our results show that stateless forwarding leads to better routing and content retrieval performance.
Vi Related Work
PIT-focused DoS attacks in CCN are a well-known problem . Rate-based [32, 33, 34] and statistical-based tests [35, 36, 37] have been proposed to detect these attacks and subsequently limit the incoming interfaces upon which malicious interests arrive. However, this only treats a symptom of the problem–it does not solve the core issue of PIT state in routers. Dai et al.  propose a technique called “interest tracebacks” to identify malicious attackers and limit the rate at which they can send messages to the network. The key observation is that PIT state leaves a trace that terminates at the source of an interest. The network can use this trail to then identify the attacker. However, this approach depends on localized attackers sending interests at a high rate; it does not work for highly distributed adversaries. Similar in-network throttling techniques were discussed in  and . Complementary to this general technique, Al-Sheikh et al.  introduce FIB exclude filters that seek to prevent malicious interests from propagating upstream to locations in the network where the target content cannot possibly be served. These filters work for static content, only, and cannot be used to prevent interests for dynamic content from being forwarded. Li et al.  propose the use of consumer-based puzzles that must be solved as a native rate-limiting technique. These puzzles, or “interest cash,” are generated by producers to be solved and must be completed for each interest. Although this approach is effective, it severely harms benign consumers.
Techniques to outright replace the PIT have also been proposed.  devised a “semi-stateful” solution wherein packets are marked (with Bloom Filters ) to be forwarded correctly. This approach shifts the state that was once in the PIT to the packets themselves and creates unnecessary communication and control overhead in the network. In a similar vein, Wang et al.  describe a protocol variant wherein resource-constrained PITs can offload the per-request state into interests that are forwarded. This technique puts PIT state “on the wire” and allows a PIT to naturally decrease in size as content is returned without dropping interests from benign consumers and routers. This is in contrast to our work where we put defer state information to the routing protocol.
Salah et al.  used a router coordination framework called CoMon (Coordination with Lightweight Monitoring) to enable adjacent nodes to share information about forwarding state and traffic. Select routers are assigned the role of “monitor.” Their goal is to monitor interest and content exchanges and measure the (un)satisfaction rate. This information is periodically reported to a central “domain controller” that is in charge of processing the traffic reports to detect and respond to IF attacks. Monitoring routers are chosen based on their location in the network and closeness to producers. This solution assumes an unrealistic static topology and centralized post facto detection mechanism. In summary, this scheme is an extension to previous rate-based throttling solutions.
Almirshari et al.  proposed a technique to “piggyback” interest and content objects to enable high throughput bidirectional communication in NDN. Their approach introduces a new packet type in addition to interests and content objects. It also requires that interests are unnaturally extended to carry application data in the name. Moreover, their approach is still susceptible to IF attacks since it requires PIT state for bidirectional communication. Dai et al.  study extensions to the PIT to support modern applications such as streaming services and online gaming. They propose to create long-lived PIT entries to enable bidirectional communication between clients and servers. This only serves to make adversary’s job easier in an IF attack.
Vii Conclusion and Future Work
Motivated by the Interest Flooding attacks in current CCN, we proposed an alternative CCN architecture without PITs, called stateless CCN. We investigated the benefits of PIT and realized that they do not significantly improve the performance of content distribution. Our proposed architecture is based on Routable Backward Names (RBNs) used to route content back towards requesting consumers. We provided a comprehensive performance and security assessment of the proposed stateless CCN architecture. We also discussed how it is practical to deploy this architecture in today’s IP networks and showed that deploying it alongside with current CCN does not achieve the expected benefits and performance.
However, removing the PIT came at the expense of losing support of CCN features and extensions developed throughout the last few years. Consumer anonymity, for instance, cannot be achieved in RBN-based stateless CCN at the network layer without using supporting protocols such as ANDāNA . Moreover, the Interest-Key Binding rule (IKB)  that allows efficient content trust at the network layer relies heavily on PIT. Thus, the IKB rule cannot be applied in RBN-based stateless CCN. We believe that the advantages of the proposed CCN architecture outweigh its drawbacks. We therefore defer solutions to the aforementioned disadvantages, e.g., anonymity and trust, to future work.
-  B. Ahlgren et al., “A survey of information-centric networking,” IEEE Communications, vol. 50, no. 7, 2012.
-  V. Jacobson et al., “Networking named content,” in CoNEXT, 2009.
-  I. Solis, “CCN 1.0 (tutorial),” in ICN, 2014.
-  C. Yi et al., “Adaptive forwarding in named data networking,” ACM CCR, vol. 42, no. 3, 2012.
-  J. Zhou et al., “A proactive transport mechanism with explicit congestion notification for NDN,” in ICC, 2015.
-  H. Park et al., “Popularity-based congestion control in named data networking,” in ICUFN, 2014.
-  Y. Wang et al., “An improved hop-by-hop interest shaper for congestion control in named data networking,” ACM CCR, vol. 43, no. 4, 2013.
-  G. Carofiglio et al., “Multipath congestion control in content-centric networks,” in INFOCOM WKSHPS, 2013.
-  S. Braun et al., “An empirical study of receiver-based aimd flow-control strategies for CCN,” in ICCCN, 2013.
-  L. Saino et al., “CCTCP: A scalable receiver-driven congestion control protocol for content centric networking,” in ICC, 2013.
-  P. Gasti et al., “DoS and DDoS in named data networking,” in ICCCN, 2013.
-  W. So et al., “Toward fast NDN software forwarding lookup engine based on hash tables,” in ANCS, 2012.
-  C. Tsilopoulos et al., “Reducing forwarding state in content-centric networks with semi-stateless forwarding,” in INFOCOM, 2014.
-  H. Yuan et al., “Scalable pending interest table design: From principles to practice,” in INFOCOM, 2014.
-  G. Carofiglio et al., “Pending interest table sizing in named data networking,” in ICN, 2015.
-  W. John et al., “Estimating routing symmetry on single links by passive flow measurements,” in IWCMC, 2010.
-  I. Fette et al., “RFC 6455: The websocket protocol,” 2011.
-  “Skype,” http://www.skype.com/.
-  B. Cohen, “The BitTorrent protocol specification,” 2008.
-  J. Burke et al., “Securing instrumented environments over content-centric networking: the case of lighting control and NDN,” in INFOCOM WKSHPS, 2013.
-  P. Gusev et al., “NDN-RTC: Real-time videoconferencing over named data networking,” in ICN, 2015.
-  Z. Zhu et al., “Let’s chronosync: Decentralized dataset state synchronization in named data networking,” in ICNP, 2013.
-  G. Carofiglio et al., “Modeling data transfer in content-centric networking,” in ITC, 2011.
-  J. Garcia-Luna-Aceves et al., “Understanding optimal caching and opportunistic caching at the edge of information-centric networks,” in ICN, 2014.
-  C. Yi et al., “A case for stateful forwarding plane,” Computer Communications, vol. 36, no. 7, 2013.
-  N. Rozhnova et al., “An extended hop-by-hop interest shaping mechanism for content-centric networking,” in GLOBECOM, 2014.
-  G. Carofiglio et al., “Joint hop-by-hop and receiver-driven interest control protocol for content-centric networks,” in SIGCOMM ICN Workshop, 2012.
-  Y. Ren et al., “An interest control protocol for named data networking based on explicit feedback,” in ANCS, 2015.
-  P. Syverson, “A taxonomy of replay attacks [cryptographic protocols],” in CSFW, 1994.
-  P. Ferguson, “RFC 2827: Network ingress filtering: Defeating denial of service attacks which employ IP source address spoofing,” 2000.
-  M. Virgilio et al., “PIT overload analysis in content centric networks,” in SIGCOMM ICN Workshop, 2013.
-  A. Compagno et al., “Poseidon: Mitigating interest flooding DDoS attacks in named data networking,” in LCN, 2013.
-  R. You et al., “Detecting and mitigating interest flooding attack in content centric networking,” Advances in Computer Science and Technology, vol. 65, 2014.
-  A. Afanasyev et al., “Interest flooding attack and countermeasures in named data networking,” in IFIP Networking, 2013.
-  N. T. Nguyen et al., “Detection of interest flooding attacks in named data networking using hypothesis testing,” 2015.
-  T. Nguyen et al., “An optimal statistical test for robust detection against interest flooding attacks in CCN,” in IFIP/IEEE IM, 2015.
-  J. Tang et al., “Identifying interest flooding in named data networking,” in GreenCom, 2013.
-  M. Mosko, “CCNx 1.0 protocol specification roadmap,” 2013.
-  M. Mosko et al., “CCNx messages in tlv format,” 2015, https://tools.ietf.org/html/draft-irtf-icnrg-ccnxmessages-00.
-  J. W. Stewart III, BGP4: inter-domain routing in the Internet. Addison-Wesley Longman, 1998.
-  M. Waldvogel, “Fast longest prefix matching: algorithms, analysis, and applications,” Ph.D. dissertation, ETH Zürich, 2007.
-  M. Kobayashi et al., “A longest prefix match search engine for multi-gigabit ip processing,” in ICC, 2000.
-  S. Dharmapurikar et al., “Longest prefix matching using bloom filters,” in SIGCOMM, 2003.
-  C. Ghali et al., “Network-layer trust in named-data networking,” SIGCOMM CCR, vol. 44, no. 5, 2014.
-  C. Ghali et al., “Needle in a haystack: Mitigating content poisoning in named-data networking,” in NDSS SENT Workshop, 2014.
-  G. Acs et al., “Cache privacy in named-data networking,” in ICDCS, 2013.
-  A. Mohaisen et al., “Protecting access privacy of cached contents in information centric networks,” in ASIA CCS, 2013.
-  S. Mastorakis et al., “ndnSIM 2.0: A new version of the NDN simulator for NS-3,” Technical Report, 2015.
-  “Network simulator 3 (NS-3),” http://www.nsnam.org/.
-  A. Afanasyev et al., “NFD developers guide,” Technical Report NDN-0021, NDN Project, Tech. Rep., 2014.
-  “DFN-Verein,” http://www.dfn.de/.
-  “DFN-Verein: DFN-NOC,” http://www.dfn.de/dienstleistungen/dfninternet/noc/.
-  H. Dai et al., “Mitigate ddos attacks in NDN by interest traceback,” in INFOCOM WKSHPS, 2013.
-  S. Al-Sheikh et al., “Revisiting countermeasures against NDN interest flooding,” in ICN, 2015.
-  Z. Li et al., “Interest cash: an application-based countermeasure against interest flooding for dynamic content in named data networking,” in CFI, 2014.
-  B. H. Bloom, “Space/time trade-offs in hash coding with allowable errors,” ACM Communications, vol. 13, no. 7, pp. 422–426, 1970.
-  K. Wang et al., “Decoupling malicious interests from pending interest table to mitigate interest flooding attacks,” in GC Wkshps, 2013.
-  H. Salah et al., “Lightweight coordinated defence against interest flooding attacks in NDN,” in INFOCOM WKSHPS, 2015.
-  M. Almishari et al., “Optimizing bi-directional low-latency communication in named data networking,” SIGCOMM CCR, vol. 44, no. 1, 2013.
-  H. Dai et al., “On pending interest table in named data networking,” in ANCS, 2012.
-  S. DiBenedetto et al., “ANDaNA: Anonymous named data networking application,” in NDSS, 2012.