TL;DR
This paper reveals that the widespread, often unfiltered use of BGP communities, intended for policy and traffic engineering, is a significant, previously overlooked cause of unnecessary BGP routing messages. By analyzing real-world BGP data and conducting controlled experiments, the authors demonstrate that these "community-only" updates can increase routing traffic by up to 25%, impacting network efficiency and storage.
What Problem Does This Paper Address?
The Border Gateway Protocol (BGP) is the backbone of the internet, responsible for routing traffic between autonomous systems (ASes). BGP communities are optional attributes that allow ASes to tag routes with meta-information for various purposes like policy enforcement, traffic engineering, and signaling. However, BGP communities lack standardized semantics, meaning each AS defines its own meaning. This has led to a practice where many ASes propagate communities they don't understand, creating potential security vulnerabilities. This paper focuses on a different, unintended consequence: the impact of this permissive propagation on the volume of BGP routing messages. Specifically, it investigates how changes in community attributes, even without changes to the actual routing path, can trigger unnecessary BGP update messages, thus increasing network overhead.
Methods
The authors employed a two-pronged approach:
- Controlled Laboratory Experiments: They set up a simulated network topology with multiple ASes and tested various BGP implementations (Cisco IOS, Juniper Junos, Nokia SR OS, BIRD, FRRouting, OpenBGPD, Quagga). These experiments were designed to isolate the impact of BGP communities on update generation under different scenarios, including the presence and absence of community filtering, and the flapping of network links. They specifically examined how changes in community attributes, with or without AS path changes, trigger updates.
- Real-World Data Analysis: They analyzed a decade of BGP routing data from RouteViews and RIPE RIS collectors. This involved examining billions of BGP announcements to quantify the prevalence of different announcement types (e.g., path changes only, community changes only, no changes) and to track trends over time. They also used "routing beacons" – specially announced and withdrawn prefixes – to better understand community exploration and its impact on update generation.
Datasets
- Real-world BGP Data:
- dhist: A dataset comprising BGP update messages collected over a ten-year span (2010-2020) from RouteViews and RIPE RIS. This dataset was used to analyze the longitudinal trends of announcement types.
- dmar20: A snapshot of the dhist dataset from March 15, 2020, providing a detailed view of recent BGP traffic.
- dbeacon: A subset of the BGP data focused on routing beacon prefixes, used to study community exploration and its impact in a more controlled manner.
- Laboratory Experiments: The authors used multiple BGP implementations in a controlled lab environment to reproduce and understand the mechanisms behind unnecessary updates.
Key Insights and Ideas
- Communities as a Silent Traffic Generator: The most striking finding is that BGP communities, even when not changing the AS path, are a significant driver of BGP update messages. This "community-only" traffic is often unnecessary.
- The "Community Exploration" Phenomenon: In scenarios like path exploration (where ASes try out different paths), the addition of diverse geolocation communities can lead to a cascade of updates, even if the underlying route hasn't fundamentally changed.
- Default Behavior is Permissive: Many BGP implementations, by default, forward communities without requiring explicit configuration. This ease of propagation contributes to the problem.
- Filtering is Not a Silver Bullet: Even when an AS filters communities on egress, some implementations still generate a duplicate update message (albeit without the community), contributing to unnecessary traffic.
- Global Withdrawals Amplify Community Information: A substantial portion of community information is revealed during global withdrawal phases, as ASes re-announce routes with different communities to explore paths.
Key Results
- Prevalence of Unnecessary Updates: In March 2020, approximately 25% of BGP announcements signaled no path change but did have a community change, and another 25% had neither path nor community change. The authors consider the former to be unnecessary.
- Traffic Reduction Potential: By removing announcements with only community changes (and no path changes), the uncompressed data volume of BGP update files could be reduced by 20-25%.
- Router Implementation Differences: While most tested router implementations forward communities by default and generate updates for community changes, Juniper Junos and Nokia SR OS showed more robust handling of duplicates.
- Community Information Leakage: Over 60% of unique community attributes were revealed exclusively during withdrawal phases of beacon prefixes, highlighting how communities can expose routing information.
Limitations and Future Work
- Limitations:
- The study primarily focuses on traditional 32-bit BGP communities, not large or extended communities, which are less common.
- While they analyzed many BGP implementations, the list is not exhaustive.
- The analysis of "unnecessary" updates is based on the assumption that community-only changes are inherently unnecessary, which might not always be the case in highly dynamic routing environments.
- Future Work:
- Develop more sophisticated network tomography techniques to classify per-AS community behavior (tagging, filtering, ignoring).
- Estimate the contribution of ASes that do not peer with collectors to the generation of unnecessary announcements.
- Explore how BGP communities can be used to understand anomalous routing behavior and predict anomalous communities.
- Investigate the implications of communities on BGP scalability and information hiding, as communities can reveal peering locations and AS relationships.
Summary updated