Multicast Musings
Recently, an interview/podcast that I recorded with Russ White and Tom Ammon on rule11.tech was published. This recording was the result of a paper that Jon Crowcroft and I self-published on arxiv, also featured in an APNIC blog, entiled If Multicast is the Answer, what was the Question?
The main intention of this paper was to make multicast interesting (again) for both the research and engineering community, thus providing. Although multicast has been around almost from the beginning of the Internet, it’s become more or less a niche technology (in its prevalent form of IP multicast), although many communication scenarios do exist where a multipoint semantic is underlying (such as, among others, the video content delivery that many of us use almost daily).
So backtracking from the answer in the form of IP multicast, we revisit some of the motivations for multicast as an extension to the unicast and broadcast semantics of the Internet in its very beginning. Efficiency is one that is often listed first, utilizing in-network replication of packets instead of relying on endpoint replication in overlay solutions. But the paper also identifies the aspect of discovery as a key motivation for multicast; replicating, as Jon puts it, the ‘watercooler moment’ that some of us may know from real life, where employees gather around the watercooler (representing a multicast group) in the morning to exchange the previous day events or upcoming future happenings.
We then tie these two aspects into the original motivation for this paper, namely a discussion in the IETF Routing Work Group interim meeting from June 2022 where Lixia Zhang posed the question on what is being identified by the identifiers? For answering this question, we argue in the paper that decoupling the WHAT is delivered from the HOW it is being delivered may be the first step in order to consolidate existing realizations for multicast with emerging new proposals. In our presented vision for refactoring communication systems around those forms of multicast realizations, we formulate a consolidated multicast semantic as
A datagram with source address S towards destinations D1, ..., Dn, which in turn are identified through destination information D, is formed as one or more responses to adequate requests from D1,..., Dn towards S, where the ephemeral channel CM is defined through an identifying characteristic across all requests from D1,..., Dn.
This consolidated semantic, which provides the suggested separation of WHAT and HOW, is then used to provide an overview of different possible answers, not limited to IP multicast only but also including a variety of information-centric networking (ICN) approaches as well as newer path-based forwarding approaches, the latter providing a strong decoupling of HOW (in the form of path identifiers) and WHAT (in the form of upper layer identifiers like URIs) is being delivered.
Why does this reconciliation matter though? Our example use cases attempt to motivate this question, including examples for multicast which many coming from the IP multicast angle would not assume to be applicable although they do exhibit a clear multipoint behaviour, particularly under our proposed consolidated semantic. One strong point for answers like ICN or path-based forwarding solutions (such as BIER or SDN-based approaches) is the support for information retrieval scenarios, most notably chunk-based video delivery as it is used in most over-the-top video platforms that many of us enjoy today. While the asynchronous and non-channel based nature of delivery in those scenarios makes the use of IP multicast cumbersome (or even impossible), the chosen delivery models of those newer approaches yield a multicast gain that could help stem the ever growing bandwidth hunger of video delivery in the Internet – find here the presentation of a demonstration I gave at the Mobile World Congress 2016, showcasing an SDN-based forwarding solution with a significant multicast gain for HTTP-based video delivery solutions.
But also newer use cases, such as distributed AI and large-scale IoT, are listed. Key to those scenarios, similar to chunk retrieval, are the highly dynamic relationships, which require new forms of multicast delivery management, such as in the form of ephemeral (source-based) path relations or interest grouping in intermediary forwarding elements (as utilized in ICN variants like named data networking).
But those use cases for multicast not only outline the opportunities, particularly for emerging solutions, but also provide a way forward to address the often cited main reason for the failure of (IP) multicast, namely the lacking economic incentives for inter-domain providers to support multicast delivery since many of the use cases, such as distributed AI or IoT but also video delivery (due to Point-of-Presence frontloading), are often limited to intra-domain delivery, thus avoiding the inter-domain issue of IP multicast.
We end our discussion in the paper by outlining the need for further research, thus returning to the original intention of our paper, namely to make multicast interesting (again) for the research community, where those research aspects span topics like addressing and routing, but also include architectural framework within which to realize possibly different forms of multicast realizations and through which to avoid the failure of widespread adoption of IP multicast (at least for the inter-domain delivery).
But we also see SDOs, like the IETF, as part of the road to make multicast interesting again. Efforts like this paper (and its advertisement on relevant IETF lists) are key to reaching this community, too. This brings me back to the Rule11 interview/podcast, which is another building block in that outreach. The great thing about the 46 minutes of chatting with both Russ and Tom was the wider coverage we could give to key aspects that surround the use of multicast in the Internet, such as the drive to increased centralization (and thus away from localized, decentralized multicast realizations of, e.g., your family chats), the issues on privacy and others you can hear us talking about. For more on the key discussions on our paper and those surrounding aspects, please listen into the podcast and enjoy!
This leaves me with giving a big thank you not just to Russ White and Tom Ammon for their chance to discuss this paper, but also to APNIC for their blog about the paper, and those who provided comments while writing the paper last year, namely Tony Li, Dino Farinacci, Lixia Zhang, and Luigi Iannone – but foremost to Jon Crowcroft for accompanying this exercise as a co-author!