SconePro with Network Jam and Clotted Streams

Last week I attended the IETF119 meeting in Brisbane (remotely), and I attended a meeting for a newly proposed working group called SCONEPRO where some internet service providers and large video content platforms want to work together to make the controversial practice of traffic shaping work slightly better. Here are my notes and thoughts. I would like to thank Mallory Knodel and Daniel Kahn Gillmor for their input and helping me make sense of all of this.

Background

The creatively named SCONEPRO (Secure Communications of of Network Properties) meeting was held on March 21, 2024 as a working-group forming BoF (Birds of a Feather) at the IETF119 Brisbane. BoF meetings like these are prequisites to setting up IETF working groups by ensuring there is enough interest within the community and that the IETF is the right place for standardization.

SCONEPRO aims to develop an internet protocol to deal with a particular use case: Network Operators, particularly mobile, often employ methods such as traffic shaping to control the flow of traffic when there is a high load on the network. This can interfere with how some applications run. SCONEPRO is particularly concerned about video applications.

Why video in particular? Not only does it form the majority of internet traffic by their estimation, video streaming applications often allow the client to adjust the bitrate (colloquially, the “quality” or “resolution” of a video), in order to reduce its impact on a congested network.

End users, through client applications, have no way of knowing for sure that their traffic is being shaped. Certain solutions exist to figure that out, but application developers argue that they are complicated and costly. At the same time, network operators usually have no way of telling what traffic is video traffic because transport encryption is so ubiquitous.

The SCONEPRO working group if established would develop a protocol that allows a network to communicate to a client application about whether it wants to do traffic shaping, and announce the bitrate that the network is willing to allow. This gives the client the option to artificially reduce the video quality on their end. They argue that this would provide a better “quality of experience” (QoE) to their users.

What Happened at the Meeting

The meeting started with a short explainer of the goal of the BoF by the chairs. I’ll give a summary of my notes and impressions, but if you’re interested to see for yourself refer to the video at this link. You can also find links to the official notes for the meeting and the slides here.

How Shapers and Policers Work

Marcus Ihlar from Ericsson gave an overview of the current state of network shaping and policing and this is my summary of that talk. There are several reasons why a network might want to throttle video, for example bandwidth limitations and congestion controls. Also, more networks are moving from a data-cap model for charging users to a bitrate-cap model, in which users can pay more to access higher resolution media.

Client applications like video streaming services often employ a technique called adaptive bitrate (ABR), where they predict the capacity of the network then dynamically change the bitrate of the video to deliver it without interuptions. Networks see this as an oppurtunity to reduce the load on their networks, so they attempt to detect when a traffic flow is video, then use traffic shapers or policers to throttle the flow artificially.

The functional difference between a shaper and a policer is that the former adds a delay to network packets to spread them out over time and policers drop packets above it’s allowed datarate policy. Traffic shapers and policers often have the same end result.

Neither technique works really well because it’s not easy to detect video content because of encryption. Network operators often employ techniques to overcome that constraint with heuristics, DPI or trying to interpret the Server Name Indicator of the unencrypted initial QUIC packet, which is not always reliable. This means that the either the shaping or the ABR might not work as planned, creating a bad user experience.

Some internet service providers have agreements with large content platforms (like Youtube) that provide video to provide traffic shaping that works more consistently but these are all proprietary.

Meta and Ericsson Experiment

Matt Joras from Meta presented the results of a feasibility study conducted by Meta and Ericsson in which they developed a SCONEPRO proof of concept. They implemented a MASQUE proxy that connected a Facebook app and a Facebook Video Content Delivery Network (CDN) server. In addition to facilitating the transfer of traffic between the CDN and the app, the proxy server also introduced a maximum send rate signal. The Facebook app and the CDN then used the send rate signal value to manually limit their bitrate to fit the self-imposed network constraint. Their takeaways was that SCONEPRO is feasible and it results in improvements to consistent video playback, but only when compared to the experience with a traffic shaper.

Lessons from History

Brian Trammell gave a presentation on the history of PLUS, a prior IETF working group where a more generalized approach to on-path network property signalling was discussed, but ultimately faltered for the following reasons. While the generalized approach was considered by many participants in the process to be good engineering, it is created various unintended dystopic consequences when you add policy considerations to those aforementioned engineering considerations. The cited example was, when engineering a header to signal loss tolerance and flow start, it was possible in some cases to infer the age of the user from these network signals.

The recommendation based on the lessons learned was to keep SCONEPRO specific and to make it optional for clients.

Discussion on Use Cases and Scope

The second half of the meeting went into discussing the use case and potential scope of a charter. Here is a my summary of key inputs as I understood them. Don’t quote anyone directly from this without reviewing the video source, any embellishments are mine.

• There were questions about the how to address network complexity, like if there are multiple shapers on the path, and the need to get the information from the box with the lowest bandwidth, which would be the actual bottleneck.
• Jason Livingood of Comcast expressed some frustration with having to revisit the discussion on traffic shaping. He mentioned that there are other solutions, such as investing in capacity, and also referenced regulatory action in the US to ban traffic shaping. Finally, networks shouldn’t sell what they can’t deliver.
• David Schinazi from Google said, “This is a case of the IETF ensuring our job security a bit longer.”
• Ted Hardie also from Google and an author of an RFC on signaling highlighted that one principle for good signal design is that there should be no incentive to fake it. He also brought up the example of the spin bit in QUIC and how IETF engineers are good at identifying side-channel attacks. Tommy Pauly from Apple expanded on that by mentioning Ted’s RFC which has additional considerations for design of path signals.
• Tom Saffell provided some insights from YouTube’s infrastructure experience and the challenges faced in implementing proprietary solutions to this problem, and said Google and YouTube are interested in working on this. YouTube are supportive of network operator efforts to reduce data tonnage. Wonho Park from Tiktok also expressed support for working on this problem, stating that traffic shaping is not optimal. There were similar supportive inputs from Suhas Nandakumar (Cisco), Jeff Smith (T-Mobile America), and Dan Druta (AT&T).
• Martin Duke expressed some concerns about the effects of this on best-effort traffic. He acknowledged the arguments that would improve best-effort by reducing incentives to do clumsy things in order to traffic shape. He also expressed concerns about extensability to other use cases.
• Lars Eggert, former chair of the QUIC working group, expressed concerns about how operators are enamored with adding complexity to manage capacity, and how that complexity is a lucaritve market for vendors. He also is worried about this being used to monetize bitrate discrimination.
• Other concerns brought up were around scalability, security, and feasibility of any possible solution, including issues related to discovery and authentication of proxies. How do you know the box giving you the signal has the authority to shape your traffic? Running so much traffic over proxies might be expensive and, ultimately, it doesn’t replace the need for shapers and policers which network operators might still use for other purposes.
• Stephen Farell, research fellow at Trinity College Dublin who studies security and networking, raised a concern about whether and how the security claims could be upheld, particularly client authentication to/of random boxes.
• Tom Saffell (YouTube) mentioned some policy considerations that should be combined with the technical solution if they were to consider implementing it, namely:
  • Transparency to users: restrictions must be visible
  • User choice: buy a plan with no restriction
  • Equal treatment: wish to be treated as any other provider
• Some comparisons were made between this and ECN (Explict Congestion Notification – RFC3168), however Matt Joras (Meta) made a point was that this is not explicitly a congestion issue, it’s an application layer signal, for example the network might be shaping traffic because of a subscriber policy.

Finally, there was a vote on whether the work group formation should move forward, 51 people voted yes, and 20 voted no, showing some opposition to this and lack of consensus.

Some Public Interest Considerations

Net Neutrality is the principle that Internet service providers must treat all communications equaly, and may not discriminate traffic based on content, particularly for profit or to disadvantage competition. Giving network operators control over bitrate, even with consent from the client, opens the door to violating net neutrality.

The fig leaf on traffic shaping is that it’s framed as a congestion control or a network capacity issue. One argument for traffic shaping has always been user choice: that users might want to prioritize a video call over updates downloading in the background. If it’s the end user’s choice as to what traffic gets shaped, and if they consent to it, then it’s no longer harmful traffic discrimination.

The problem remains that we have to take the network operators’ word that these techniques are only applied when congestion happens, and not to extract more profit, or push users into paying more for higher bitrates artificially. SCONEPRO offers a “QoE” improvement over the status quo in (physically or artificially) capacity constrained networks, but user “QoE” would also improve if the capacity of the network is increased. In the case of a protocol that requires opt-in from the application, this can lead to business partnerships that create a “fast lane,” which is another common net neturality violation.

SCONEPRO currently proposes some design goals in the proposed charter that might be relevant to these issues:

1. Associativity with an application. The network properties must be associated with a given application traversing the network, for example a video playback.
2. Client initiation. The communication channel is initiated by a client device.
3. Network properties sent from the network. The network provides the properties to the client. The client might communicate with the network, but won’t be providing network properties.
4. On-path establishment. That is, no off-path element is needed to establish the communication channel between the entity communicating the properties and the client.
5. Optionality. The communication channel is strictly optional for the functioning of application flows. A client’s application flow must function even if the client does not establish the channel.
6. Properties are not directives. A client is not mandated to act on properties received from the network, and the network is not mandated to act in conformance with the properties.
(…)
9. Security. The mechanism must ensure the confidentiality, integrity, and authenticity of the communication. The mechanism must have an independent security context from the application’s security context.

SCONEPRO is being framed as a solution to improve user experience, however most of the proponents seem to be telecom providers and major content platforms. I think SCONEPRO is a marginal improvement over the status quo in which traffic shaping is achieved with proprietary solutions and agreements between telecoms and major platforms.

One important consideration would be the effects of SCONEPRO deployment in different regulatory enviroments. In places where net neutrality protections are not robust, providing a “bitrate signal” or future signals based on use cases invented in the future may enable profited-based traffic discrimination.

It’s not clear how some of the desired properties of SCONEPRO such as optionality or not being a directive can be technically enforced, which means when looking at the effects of introducing such a protocol these design goals can be safely ignored. Client applications that implement SCONEPRO gain an advantage over those who don’t even if all the rules are respected, and if not, this opens the door to enable telecoms to more easily offer tiered services, zero-rating and fast lanes.

Ultimately, I do agree with the BoF’s premise that there is a problem to be solved but it’s not by encoding the status quo into the protocols of the internet. I think the practice of content-based traffic shaping needs to be better looked at and tackled from a regulatory and consumer advocacy standpoint. ABR traffic shaping and by extension SCONEPRO takes choice away from users and negogiates application parameters on the network in an opaque way to force data austerity on them.

Did you find this helpful or have some feedback? Would you like to see a follow up dive into similar prior work at the IETF like PLUS, MINUS, SPUD, or SADCDN? Reach out and let me know.