IRR & Routing Policy Configuration Considerations

The purpose of this document is to catalog past issues influencing the efficacy of Internet Routing Registries (IRR) for inter-domain routing policy specification and application in the global routing system over the past two decades. Additionally, it provides a discussion regarding which of these issues still pose problems in practice, and which are no longer obstacles, but whose perceived drawbacks continue to stifle inter-provider policy-based filtering support and IRR utility to this day.

IRRs can be used to express a multitude of Internet number bindings and policy objectives, to include bindings between an origin AS and a given prefix, AS and community import and export policies for a given AS, as well as AS macros (as-sets in RPSL-speak) that convey the set of ASes for which a given AS intends to include in some common group. As quoted from Section 7 of , Routing in a Multi-Provider Internet: While ensuring Internet-wide coordination may be more and more difficult, as the Internet continues to grow, stability and consistency of the Internet-wide routing could significantly benefit if the information about routing requirements of various organizations could be shared across organizational boundaries. Such information could be used in a wide variety of situations ranging from troubleshooting to detecting and eliminating conflicting routing requirements. The scale of the Internet implies that the information should be distributed. Work is currently underway to establish depositories of this information (Routing Registries), as well as to develop tools that analyze, as well as utilize this information.

The term IRR is often used, incorrectly, as a broad catch-all term to categorize issues related to the accuracy of data in the IRR, the Routing Policy Specification Language (RPSL) and the operational deployment of policy (derived from RPSL contained within the IRR) to routers. It is important to classify these issues into distinct categories so that the reader can understand which categories of issues are historical artifacts that are no longer applicable and which categories of issues still exist and might be addressed by the IETF. The following sections will separate out challenges related to the IRR into the following categories. First, accuracy and integrity of data contained within the IRR. Second, the Resource Policy Specification Language (RPSL) used to represent various types of data in the IRR. Third, operation of the IRR infrastructure, i.e.: synchronization of resources from one IRR to other IRRs. Finally, the methods related to extraction of policy from the IRR and the input plus activation of that policy within routers.

The following section will examine issues related to accuracy and integrity of data contained within the IRR.

Internet number resources include IPv4 addresses, IPv6 addresses, Autonomous System Numbers (ASNs), and more. While these resources are generally allocated by hierarchical authorities, a general mechanism for formally verifying (such as through cryptographic mechanisms) when parties have been allocated resource remains an open challenge. We generally define such a system a Resource Certification System, and we note that some candidate examples of how such a general system might be implemented and deployed exist , , . One of the largest weaknesses often cited with the IRR system is that the data contained within the IRRs is out of date or lacks integrity. This is largely attributable to the fact that existing IRR mechanisms do not provide ways for a relying party to (cryptographically) verify the validity of an IRR object. That is, there has never existed a resource certification infrastructure that enables a resource holder to authorize a particular autonomous system to originate network layer reachability advertisements for a given IPv4 or IPv6 prefix. It should be noted that this is not a weakness of the underlying Routing Policy Specification Language (RPSL) , but rather, was largely the result of no clear demand by the operator community for Internet Number Resource Registries to provide sufficient resource certification infrastructure that would enable a resource holder to develop a cryptographic binding between, for example, a given AS number and an IP prefix. Another noteworthy (but slightly different) deficiency in the IRR system is the absence of a tangible tie between the resource and the resource holder. That is, generally there is no assurance of the validity of objects at their creation time (except for a subset of, for example, the RIPE IRR where RPSS attests for RIPE address holders and RIPE ASN holders). If a resource holder's authorization cannot be certified, then consumers cannot verify attestations made. In effect, without resource certification, consumers are basically only certifying the assertions that the creator / maintainer of the resource object has made (not if that assertion is valid). The RIPE community addressed this last issue by putting in a foundation policy , which requires a contractual link between the RIPE NCC and the end user in direct assignment + ASN assignment cases, which weren't previously covered by LIR contracts for address allocations. There were a couple of intentions with this policy: There was an encumbrance placed in the policy for the LIR to charge the end-user for provider independent resources. This action created a collection mechanism for PI address resources (IPv4 / IPv6 space, ASNs). It guaranteed that all RIPE NCC allocated/assigned space would be subject to a contractual link, and that this contractual chain might end up actually meaning something when it came to the issue of who made what claim about what number resource. It tied into the RIPE NCC's object grandfathering policy which ties the registration details of the end-user to the object registered in the IRR database. While this policy specifically addressed PI/portable space holders, other regions address this issue, too. Further, it is indeed a prerequisite for resource certification, though it does not directly address the IRR deficiencies. One of the central observations of this policy was that without a chain-of-ownership functionality in IRR databases, the discussion of certifying their contents becomes moot.

A second problem with data contained in the IRRs is that the incentives for resource holders to maintain both accurate and up-to-date information in one, or more, IRRs; including deletion of out-of-date or stale data from the IRRs can diminish rapidly when changing their peering policies (such as switching transit providers). Specifically, there is a very strong incentive for an ISP's customers to register new routing information in the IRR, because some ISPs enforce a strict policy that they will only build or update a customer's prefix-lists applied to the customer's inbound eBGP sessions based off information found within the IRRs. Unfortunately, there is little incentive for an ISP's customers to remove out-of-date information from an IRR, most likely attributed to the fact that some ISPs do not use, or enforce use of, data contained within the IRRs to automatically build incoming policy applied to customer's eBGP sessions. For example, if a customer is terminating service from one ISP that requires use of IRR data to build incoming policy and, at the same time, enabling service with another ISP that does not require use of IRR data, then that customer will likely let the data in the IRR become stale or inaccurate. Further, policy filters are almost exclusively generated based on the origin AS information contained within IRR route objects and used by providers to filter downstream transit customers. Since providers only look for route objects containing the origin AS of their downstream customer(s), stale route objects with historical and, possibly, incorrect origin AS information are ignored. Assuming that the downstream customer(s) do not continue to announce those routes with historical, or incorrect, origin AS information in BGP to the upstream provider, there is no significant incentive to remove them as they do not impact offline policy filter generation nor routing on the Internet. On the other hand, the main incentive that causes the Service Provider to work with downstream customer(s) is when the resultant filter list becomes too large that it is difficult for it to be stored on-board PE routers; however, this is more practically an operational issue with very old, legacy PE routers, not more modern PE router hardware with more robust Control Plane Engines.

A third challenge with data contained in IRRs is that it is not possible for IRR operators, and ISPs who use them, to proactively remove (perceived) out-of-date or "stale" resources in an IRR, on behalf of resource holders who may not be diligent in maintaining this information themselves. The reason is that, according to the Resource Policy Specification Language (RPSL) , only the resource holder ('mntner') specified in a 'mnt-by' value field of an RPSL resource is authorized to add, modify or delete their own resources within the IRR. To address this issue, the 'auth-override' mechanism was later developed that would have enabled a third party to update and/or remove "stale" resources from the IRR. While it is unclear if this was ever implemented or deployed, it does provide language semantics needed to overcome this obstacle. Nevertheless, with such a mechanism in place, there is still a risk that the original RPSL resource holder would not receive notifications (via the 'notify' attribute in various RPSL resources) about the pending or actual removal of a resource from the IRR in time to halt that action if those resources were still being used. In this case, if the removal of a resource was not suspended, it could potentially result in an unintentional denial of service for the RPSL resource holder when, for example, ISPs automatically generated and deployed a new policy based on the newly removed resources from the IRR, thus dropping any reachability announcement for the removed resource in eBGP.

According to , within an RPSL resource "the source attribute specifies the registry where the object is registered." Note that this source attribute only exists within the RPSL resource itself. Unfortunately, given a specific resource, (e.g.: a specific IPv4 or IPv6 prefix), most of the time it is impossible to tell a priori the authoritative IRR where to query and retrieve an authoritative copy of that resource. This makes it difficult for consumers of data from the IRR to automatically know the authoritative IRR of a resource holder, which will contain their most up-to-date set of resources. This is typically not a problem for an ISP that has a direct (customer) relationship with the resource holder, because the ISP will ask the resource holder which (authoritative) IRR to pull their resources from on, for example, a "Customer BGP Order Form". However, third parties that do not have a direct relationship with the resource holder, have a difficult time attempting to infer the authoritative IRR, preferred by the resource holder, that likely contains the most up-to-date set of resources. As a result, it would be helpful for third parties if there was a robust referral mechanism so that a query to one IRR would be automatically redirected toward the authoritative IRR for the most up-to-date and authoritative copy of that particular resource. This problem is worked around by individual IRR operators storing a local copy of other IRR's resources, through periodic mirroring, which allows the individual IRR to respond to a client's query with all registered instances of a particular IRR resource that exist in both the local IRR and all other IRRs. Of course, the problem with this approach is that an individual IRR operator is under no obligation to mirror all other IRRs and, in practice, some IRRs do not mirror the resources from all other IRRs. This could lead to the false impression that a particular resource is not registered or maintained at a particular IRR. Furthermore, the authentication process of accepting updates by any given IRR may, or may not, be robust enough to overcome impersonation attacks. As a result, there is no rigorous assurance that a mirrored RPSL statement was actually made by the authorized resource holder.

There are no provisions in the IRR mode for ensuring the confidentiality component for clients issuing queries. The overall Confidentiality, Integrity, and Availability (CIA) model of the system does lack this component, because the interface to IRRs is over an unencrypted TCP connection to port 43. This leaves the transaction open to inspection such that an adversary could be able to inspect the query and the response. However, the IRR system is intended to be composed of public policy information, and protection of queries was not part of the protection calculus when it was designed, though the use of Transport Layer Security (TLS) would address protections of query information.

All of the aforementioned issues related to integrity and accuracy of data within the IRR stem from a distinct lack of resource certification for resources contained within the IRR. Only now is an experimental test bed that reports to provide this function (under the auspices of the Resource PKI ) being formally discussed, which could also aid in certification of resources within the IRR. It should be noted that the RPKI is only currently able to support signing of Route Origin Authorization (ROA) resources that are the equivalent of 'route' resources in the IRR. There has been some sentiment that the RPKI currently is not scoped to address the same set of issues and the nuanced policy applications that providers leverage in RPSL. It is unclear if, in the future, the RPKI will be extended to support additional resources that already exist in the IRR, e.g.: aut-num, as-net, route-set, etc. Finally, a seemingly equivalent resource certification specification for all resources in the IRR has already been developed , however it is unclear how widely it was ever implemented or deployed.

Currently, several IRRs use a Near-Real-Time Mirroring (NRTM) protocol to replicate each others contents. However, this protocol has a several weaknesses. Namely, there is no way to validate that the copy of mirrored source is correct and synchronization issues have often resulted. Furthermore, the NRTM protocol does not employ any security mechanisms. The NRTM protocol relies on a pull mechanism and is generally configured with a poll interval of 5 to 10 minutes. There is currently no mechanism to notify an IRR when an update has occurred in a mirrored IRR so that an immediate update can be made. Some providers employ a process of mirroring an instance of an IRR that involves downloading a flat text file copy of the IRR that is made available via FTP . These FTP files are exported at regular intervals of typically anywhere between 2 and 24 hours by the IRRs. When a provider fetches those text files, it will process them to identify any updates and reflect changes within its own internally maintained database. The use of an internally maintained database is out-of-scope of this document, but is generally used to assist with more straightforward access to or modification of data by the IRR operator. Providers typically employ a 24 hour cycle to pull updated resources from IRRs. Thus, depending on when resource holders submitted their changes to an IRR, it may take up to 24 hours for those changes to be reflected in their policy configurations. In practice, it appears that the RPKI will also employ a 24 hour cycle whereby changes in resources are pushed out to other RPKI caches . IRRs originated from Section 7 of , specifically: "The scale of the Internet implies that the [routing requirements] information should be distributed." Regardless, the practical effect of an organization maintaining its own local cache of IRR resources is an increase in resource resiliency (due to multiple copies of the same resource being geographically distributed), a reduction in query time for resources and, likely, a reduction in inter-domain bandwidth consumption and associated costs. This is particularly beneficial when, for example, an ISP is evaluating resources in an IRR just to determine if there are any modifications to those resources that will ultimately be reflected in a new routing policy that will get propagated to (edge) routers in the ISP's network. Cache locality results in reduced inter-domain bandwidth utilization for each round trip. On the other hand, it is unclear from where the current provider replication interval of 24 hours originated or even whether it still provides enough freshness in the face of available resources, particularly in today's environment where a typical IRR system consists of a (multi-core) multi-GHz CPU connected to a network via a physical connection of 100 Mbps or, more likely, higher bandwidth. In addition, increasing bandwidth demands have shifted demand towards circuit sizes used by ISPs to 10 Gbps, thus eliminating bandwidth as a significant factor in the transfer of data between IRRs. Furthermore, it should be noted that Merit's Internet Routing Registry Daemon (IRRd) , uses 10 minutes as its default for "irr_mirror_interval". Lastly, it should be noted that , Routing Policy System Replication, attempted to offer a more methodical solution for distributed replication of resources between IRRs. It is unclear why this RFC failed to gain traction, but it is suspected that this was due to that specification's reliance on , Routing Policy System Security, that addressed "the need to assure integrity of the data by providing an authentication and authorization model." Indeed, attempts to add an otherwise absent security model to the integrity of policy statements made in RPSL. Without formal protections, it is possible for anyone to author a policy statement about an arbitrary set of resources, and publish it (as discussed above in Section .

Ultimately, the length of time it takes to replicate resources among IRRs is, generally, the dominant factor in reflecting changes to resources in policy that is eventually applied within the Control Plane of routers. The length of time to update network elements will vary considerably depending on the size of the ISP and the number of IRR resources that were updated during any given interval. However, there are a variety of common techniques, that are outside the scope of this document, that allow for this automated process to be optimized to considerably reduce the length of time it takes to update policies in the ISP's network. An ISP will begin the process of updating the policy in their network, by first fetching IRR resources associated with, for example, a customer ASN attached to their network. Next, the ISP constructs a new policy associated to that customer and then evaluates if that new policy is different from existing policy associated with that same customer. If there are no changes between the new and existing policy associated with that customer, then the ISP does not make any changes to the policy in their routers specific to that customer. On the other hand, if the new policy does reflect changes from the existing policy for that customer, then the ISP begins a process of uploading the new policy to the routers attached to that customer. The process of constructing a new policy involves use of a set of programs, e.g.: IRRtoolset, that performs recursive expansion of an RPSL aut-num resource, that comprises an arbitrary combination of other RPSL aut-num, as-set, route and route-set resources, according to procedures defined by RPSL. The end result of this process is, traditionally, a router vendor dependent configuration snippet that defines the routing policy for that customer. This routing policy may consist of the set of IPv4 or IPv6 prefixes, associated prefix lengths and AS_PATH's that are supposed to be accepted from a customer's eBGP session. However, if indicated in the appropriate RPSL resource, the policy may also set certain BGP Attributes, such as MED, AS_PATH prepend value, LOCAL_PREF, etc., at either the incoming eBGP session from the customer or on static routes that are originated by the resource holder. An ISP's customers may not adequately plan for pre-planned maintenance or, more likely, need to rapidly begin announcing a new IP prefix as a result of, for example, an emergency turn-up of the ISP customer's new downstream customer. Unfortunately, the routine, automated process employed by the ISP means that they may not begin updating their routing policy on their network for up to 24 hours, because the ISP or the IRRs the ISP uses might only mirror changes to IRR resources once every 24 hours. The time interval for the routine/automated process is not responsive to the needs of directly paying customer(s) who need rapid changes in their policy in rare situations. In these situations, when a customer has an urgent need for updates to take effect immediately, they will call the Network Operations Center (NOC) of their ISP and request that the ISP immediately fetch new IRR objects and push those changes out to their network. This is often accomplished in as little as five minutes from the time a customer contacts their ISP's NOC to the time a new filtering policy is pushed out to the PE routers that are attached to that customer's Attachment Circuits (AC's). It is conceivable that some ISPs automate this using out-of-band mechanisms as well, although the authors are unaware of any existing mechanisms that support this. Ultimately, the aforementioned latency with respect to "emergency changes" in IRR resources that need to be reflected in near-real-time in the network is compounded if the IRR resources were being used by third party ISPs to perform filtering on their peering circuits, where typically no such policies are employed today for this very reason. It is likely that the length of time that it takes IRRs to mirror changes will have to be dramatically reduced. There will need to be a corresponding reduction in the time needed by ISPs to evaluate whether those changes need to be re-compiled and reflected in router policies, which would then get pushed out to ASBR's and connected to peering circuits on their network.

As mentioned previously, after a resource holder made changes to their resources in an IRR, those changes would automatically be distributed to other IRRs, ISPs would then learn of those changes, generate a new BGP policies, and then apply those to the appropriate subset of routers in their ASes. However, in the past, one additional step is necessary in order to have any of those new BGP policies take effect in the Control Plane and to allow/deny the updated resource from a customer to their ISP and from their immediately upstream ISP to the ISP's peers. It was necessary (often manually) to actually induce BGP on each router to apply the new policy within the Control Plane, thus learning of a newly announced/changed IP prefix (or, dropping a deleted IP prefix). Unfortunately, most of these methods were not only highly operationally impactful, but also affected traffic forwarding to IP destinations that were unrelated to the most recent changes to the BGP policy. Historically, a customer would have to (re-)announce the new IP prefix toward their ISP, but only after the ISP had placed the new BGP policies into affect. Alternatively, the ISP would have to reset the entire PE-CE eBGP session either by: a) bouncing the entire interface toward the customer, (e.g.: shutdown/no shutdown); or, b) clearing the eBGP session toward the customer, (e.g.: clear ip bgp neighbor >IP address of CE router<). The latter two cases were, of course, the most highly impactful and thus could generally only be performed off-hours during a maintenance window. Once the new IP prefix has been successfully announced by the customer and permitted by the newly updated policy at the ISP's PE's (attached to that customer), it would be propagated to that ISP's ASBR's, attached to peers, at the perimeter of the ISP network. Unfortunately, if those peers had either not yet learned of the changes to resources in the IRR or pushed out new BGP policies (and, reset their BGP sessions immediately afterward) incorporating those changes, then the newly announced route would also get dropped at the ingress ASBR's of the peers. Ultimately, either of the two scenarios above further lengthen the effective time it would take for changes in IRR resources to take effect within BGP in the network. Fortunately, BGP has been enhanced over the last several years in order that changes within BGP policy will take affect without requiring a service impacting reset of BGP sessions. Specifically, BGP soft-reconfiguration ( Section 1) and, later, Route Refresh Capability for BGP-4 were developed so that ISPs, or their customers, could induce BGP to apply a new policy while leaving both the existing eBGP session active as well as (unaffected) routes active in both the Loc-RIB and, more importantly, FIB of the router. Thus, using either of these mechanisms, an ISP or its peers currently will deploy a newly generated BGP policy, based on changes to resources within the IRR, and immediately trigger a non-service impacting notification to the BGP process to have those changes take effect in their Control Plane, either allowing a new IP prefix to be announced or an old IP prefix to be dropped. This dramatically reduces the length of time after changes are propagated throughout the IRRs to when those changes are actually operational within BGP policy in an ISP's network.

Routers in the mid 1990's rarely supported incrementally updatable prefix filters for BGP, and therefore, if new information was received from a policy or internal configuration database that would impact a policy applied to a given eBGP peer, the entire prefix list or access list would need to be deleted and rewritten, compiled, and installed. This was very tedious and commonly led to leaked routes during the time when the policy was being rewritten, compiled, and applied on a router. Furthermore, application of a new policy would not automatically result in new ingress or egress reachability advertisements, from that new policy, because routers at the time would require a reset of the eBGP sessions for routing information to be evaluated by the new policy. Of course, resetting of an eBGP session had implications on traffic forwarding during the time the eBGP session was reestablished and new routing information was learned. Routers now support the ability to perform incremental, and in situ, updates to filter lists consisting of IP prefixes and/or AS_PATH's that are used within an ingress or egress BGP policy. In addition, routers also can apply those incremental updates to policy, with no traffic disruption, using BGP soft-reconfiguration or BGP Route Refresh, as discussed in the previous section.

Historically, routers had very limited storage capacity and would have difficulty in storing an extremely large BGP policy on-board. This was typically the result of router hardware vendors using an extremely limited amount of NVRAM for storage of router configurations. Another challenge with historical router hardware was that writing to NVRAM was extremely slow. For example, when the router configuration had changed as a result of updating a BGP policy that needed to accommodate changes in IRR resources, this would result in extremely long times to write out these configuration changes. Sometimes, due to bugs, this would result in loss of protocol keep-alives. This would cause an impact to routing or forwarding of packets through the platform. The above limitations have largely been resolved with more modern equipment from the last few years shipping with increasing amounts of non-volatile storage such as: PCMCIA or USB Flash Cards, Hard Disk Drives or Solid State Disk Drives. However, as capacities and technologies have evolved on modern routing hardware, so have some of the scaling requirements of the data. In some large networks, configuration growth has begun to "pose challenges" . While the enhancements of hardware have overcome some historical limitations, evidence suggests that further optimizations in configuration processing may be needed in some cases. Some of the more recent operational issues include scheduler slips and protracted commit times. This suggests that even though many historical hurdles have been overcome, there are still motivations to optimize and modernize IRR technologies.

Historically, there has not been a standardized network configuration modeling language or an associated method to update router configurations. When an ISP detected a change in resources within the IRR, it would fashion a router vendor dependent BGP policy and upload that to the router usually via the following method. First, an updated BGP policy configuration 'snippet' is generated via processes running on an out-of-band server. Next, the operator uses either telnet or SSH to login to the CLI of a target router and issue router vendor dependent CLI commands that will trigger the target router to fetch the new configuration snippet via TFTP, FTP or Secure Copy (SCP) stored on the out-of-band server. The target router will then perform syntax checking on that configuration snippet and, if that passes, merge that configuration snippet into the running configuration of the router's Control Software. At this point, the new BGP policy configuration 'snippet' is active within the Control Plane of the router. One last step remains, which is the operator will issue a CLI command to induce the router to perform a "soft reset", via BGP soft-reconfiguration or BGP Route Refresh, of the associated BGP session in order to trigger the router to apply the new policy to routes learned from that BGP session without disrupting traffic forwarding. More recently, operators have the ability to use NETCONF / SSH (or, TLS) from an out-of-band server to push a BGP configuration 'snippet' from an out-of-band server toward a target router that has that capability. However, this activity is still dependent on generating, via the out-of-band server, a router vendor dependent XML configuration snippet that would get uploaded via SSH or TLS to the target router. In the future, the ability to upload new Route Origin Authorization (ROA) information may be provided from the RPKI to routers via the RPKI-RTR protocol. However, this will not allow operators the ability to upload other configuration information such as BGP policy information, such as AS_PATH's, BGP communities, etc. that might be associated with that ROA information, like they can from IRR generated BGP policies.

As discussed above, many of the problems that have traditionally stifled IRR deployment have, themselves, become historical. However, there are still real operational considerations that limit IRR usage from realizing its full effectiveness. The potential for IRRs to express inter-domain routing policy, and to allow relying parties to learn policy, has always positioned it as a strong candidate to aid the security postures of operators. However, while routing density and complexity have grown, so have some of the challenges facing IRRs (even today). Because of this state increase, the potential to model all policies for all ASes in all routers may still remain illusive. In addition, without an operationally deployed resource certification framework that can tie policies to resource holders, there is a fundamental limitation that still exists.

One of the central concerns with IRRs is the ability of an IRR operator to remotely influence the routing operations of an external consumer. Specifically, if the processing of IRR contents can become burdensome, or if the policy statements can be crafted to introduce routing problems or anomalies, then operators may want to be circumspect about ingesting contents from external parties. A resource certification framework should be used to address the authorization of IRR statements to make attestations and to assertions (as mentioned in , and discussed above in Section ). Additionally, the external and systemic dependencies introduced by IRRs and other such systems employed to inform routing policy, and how tightly or loosely coupled those systems are to the global routing system and operational networks, introduces additional vectors that operators and system architects should consider when evaluating attack surface and service dependencies associated with those elements. These attributes and concerns are certainly not unique to IRRs, and operators should evaluate the implications of external systems and varying degrees of coupling and operational buffers that might be installed in their environments.

There are no IANA considerations.

The editors would like to acknowledge and thank Chris Morrow, Jeff Haas, Wes George, and John Curran for their help and constructive feedback.