On (De-)Centralized Communications: Part 4

#Layer 8

Part 1, Part 2 and Part 3 were technical in nature, examining communications networks. However, we also have to consider the so-called “layer 8”, the human part, which (to be more specific) spans three additional layers atop the standard ISO/OSI model, the layers 8 to 10.

Bruce Schneier and RSA define the “missing” layers from the standardized 7-layer models as follows:

• Layer 8: The individual person.
• Layer 9: The organization.
• Layer 10: Government or legal compliance

There are ongoing discussions on what centralization actually means in all kinds of spaces. One of the more interesting trends is that people question whether the term “centralization” is actually a good one to use. What the previous parts were intended to demonstrate is that yes, actually, it is!

The reason it is lies in its utility: we have available mathematical definitions of “centrality” that apply to network/graph theory. So when considering network architecture, there is little ambiguity in those terms. The prior parts connected this to Baran’s related terms of “decentralized” and “distributed” systems, to illustrate that also those terms fit the mathematical underpinnings.

There are, however, two components to the model that are often overlooked, which the previous parts explored:

• On the one hand, the relevant degree of centralization depends on the node(s) one considers, and
• on the other hand, the choice of relevant nodes depends on the layer one considers.

The last part of this series started introducing “human layer” notions, but it’s important to understand that that layer was still technical in nature. It was referring to the technical proxy for individual humans, the “user agent”, and how it is affected by architectural choices.

This is a different kind of “human” layer than Schneier’s “layer 8” refers to. In this expanded OSI model, layers 8 to 10 refer to social (and perhaps neurological) constraints on architecture.

To summarize Schneier briefly, he discusses user experience design at layer 8, organizational policy design at layer 9, and legal compliance constraints at layer 10. Of course, his writing is about security – but the same distinctions can be extended to other fields.

The key issue here is that all three additional layers influence technical design choices, and technical constraint influence layer 8-10 choices.

One example that keeps coming up is that law enforcement and politicians ask for “safe” backdoors to encryption schemes that they can use to help protect people. Unfortunately, mathematics does not distinguish between “safe” and “unsafe”, so a backdoor in an encryption scheme is equally exploitable by bad and good actors. This is layer 10 trying to shape technical layers, and the technical layer being unable to comply.

#Terminology and Taxonomy Trends

With something comparable to these additional three layers in mind, some folk argue that “centralization” is a bad term for the issues we’re facing. The counter proposal tends to be “consolidation”.

The motivation behind this attempt is good: it acknowledges that layers above the technical ones may have a more important impact than technical choices.

To provide examples of this, the DINRG group currently considers the A Taxonomy of Internet Consolidation¹ draft, as was mentioned in part 1. This draft in turn references Mark Nottingham’s Centralization, Decentralization, and Internet Standards², which already introduces the term “consolidation” as part of its informal terminology. Mark McFadden merely tries to formalize this terminology, and from that formalization derive a more useful taxonomy.

McFadden focuses on economics in his draft, convincingly arguing that market forces influence the consolidation of e.g. network infrastructure into the hands of fewer organizations. And he’s entirely correct that this carries risks with it that are independent of technical layers.

It’s perhaps also worth pointing out that this focus on economics is also a little limited. More broadly speaking, economic influence is merely one of many forms of power; it may therefore be better to address power more directly in those documents.

Once you do that, it becomes immediately obvious that the motivation behind the “consolidation” terminology is actually a layering problem. We only need to discuss the value of this term to the conversation if we fail to acknowledge that our expanded 10-layer model already has a place for networks of power – or what social theories might more often term “power structures”:

• layer 8, the individual layer, can be seen as being about the networks of mutual influence that interpersonal relationships create. This can also refer to the networks of power within an organization.
• layer 9, the organization layer, is more concerned with the networks of partnerships and service contracts that create networks of power between different organizations.
• layer 10, the government layer, represents the political networks of power that international treaties and supranational organizations such as e.g. NATO produces.

This view focuses on politics. It can also be applied to economics:

• layer 8, the individual layer, in economics is about the purchasing choices individuals make based on the availability and pricing of goods and services, in this case: networking related goods and services.
• layer 9, the organization layer, is about the strategic economic choices organizations make in order to best optimize their cost-effectiveness.
• layer 10, the (renamed) market layer, refers to the overall market and how one organization’s choice influences the others – but also about how government policy can shape the market to make some choices more attractive than others.

Armed with this conceptual model, we can go back to the network theory definitions of centrality. As you may recall, the previous parts tried to illustrate that for many of our commonly held assumptions on what centrality means, the definition of “betweenness centrality” is often the best.

Applied to layers 8-10, “betweenness” is precisely the same as “power”: it reflects on the question how much an individual, an organization or a government influences how two other entities at the same layer interact: if nobody can broker change within the European Union’s policies until either of France or Germany weigh in, they have high “betweenness centrality” within the network of power inside the EU. If one cannot get a route between two continents established without traversing the network of one core network provider, it has high “betweenness centrality” in the network at layer 9.

Consolidation isn’t a bad term. But it is also just an alias for centralization, which focuses on layers 8-10.

#Effects

Replacing the term “centralization” with this highly important “consolidation” term, however, also has some negative side effects.

First off, it effectively shifts blame.

I do not think that is the intended effect, for what it’s worth. But if conversations, for good reasons, start debating consolidation rather than centralization, it implies that all of the efforts to effect change need to be focused on those additional upper layers. This is also reflected in the often quoted criticism that technology oriented people try to solve social issues with more technology, and then are surprised when that doesn’t work.

But what we have seen with the encryption example above is that each layer’s constraints influence the others. It is correct to say that technology cannot solve social problems. But it is also incorrect to claim that technology has no influence on the shape of society. The Internet, and how it turned the world into a global village (even if unevenly distributed) is the best example of the power technology has³.

I worry that shifting the conversation from centralization to consolidation is roughly equivalent to shifting the overton window. Rather than focusing on the technological choices tech people can make, it instead absolves tech folk of responsibility, because the it strengthens the perception that real problem is power, actually.

The second effect it has is that it fails to provide the clarity it intends to bring. But for that, we have to analyse terminology a little more.

#Terminology Analysis

This in some sense refers to section 2.1 of McFadden’s draft, where he defines the terms “centralization”, “consolidation” and “concentration”, keeping in mind that the focus he has is economic power.

• Consolidation is the process of limiting choices.
• Centralization is the process of removing the power to execute change from the edge to the core.
• Concentration is the market effect of having a small number of players either dominate or monopolize a marketplace.

There are somewhat unrelated comments that probably ought to be made on this:

1. The definition of “centralization” here is very close to the definition of “betweenness centrality” in networks of power: a node with high betweenness does remove or limit power to execute change, but it’s more about the end-to-end direction than from the edge to the core.
2. The definitions of “consolidation” and “centralization” here are conflicting with dictionary definitions!

That latter point should be supported by evidence.

Centralization:

1. the act or fact of centralizing; fact of being centralized.
2. the concentration of administrative power in a central government, authority, etc.
3. Chiefly Sociology.
   1. a process whereby social groups and institutions become increasingly dependent on a central group or institution.
   2. concentration of control or power in a few individuals.

Consolidation (shortened):

1. an act or instance of combining or consolidating into a single or unified whole; the state of being consolidated; unification (…)
2. solidification; strengthening: consolidation of principles and beliefs.
3. something resulting from a unification of two or more elements; a consolidated whole.
4. Law. the union of two or more claims or actions at law for trial or appeal.
5. Finance. debt consolidation.
6. Business, Finance.
   1. the combining of several financial accounts into a single one, as when the financial results from two or more businesses are combined into a single statement.
   2. a statutory combination of two or more corporations.

(…)

The only useful definition of “concentration” the site provides, however, is “something concentrated”, which tech folk might recognize as a recursive definition. The etymology of the verb "to concentrate" is more useful, describing it as “to bring or come to a common center”.

In terms of this presented terminology, then:

• “Centralization” already is commonly defined in terms that closely match the previously formulated notion of “betweenness centrality” in networks of power. Transferring this notion of centralization also to communications is not wrong, considering that power and communications rely on each other and follow pathways at all of the above layers.
• “Consolidation” refers to something else entirely, namely the unification of several things.
• “Concentration” is largely synonymous with “centralization”, and doesn’t seem to add much value.

However, the term “consolidation” can now also be more explicitly understood as being related to layering. And I think that continues to make it useful.

Consider several network (layer 1-7) paths between two nodes A and B, going through C1 to Cn. If the majority of that latter set of nodes is controlled by a single individual, organization or government, the power to affect layer 1-7 communications is centralized in that entity – and so can be treated as a unified, single node in layers 8-10.

In other words, the dictionary definitions of “centralization” and “consolidation” seem to confirm the notions presented so far, that “centralization” as “betweenness centrality” works well across all ten layers – and that “consolidation” is effectively an alias of “centralization” when applied to the top three.

#Other Considerations

The idea that it’s worthwhile to separate out different layers in the “centralization” problem on the one hand, but map influences between the layers on the other, is also shared by the authors of Proliferation of the Service-centric Distributed Consensus Model and its Impact on Ethereum⁴.

To quote David Guzman from the DINRG mailing list:

There is a layered discussion behind Internet consolidation.

Governance and Economics may be an Exogenous Space (two screwdrivers), Name Resolution is another space, Consensus is a Participation Space, and messaging (e.g., peer-to-peer) is a Communications Space. These spaces are supported by a Resource Execution Space (infrastructure). For example, anyone can provide an execution environment.

With the last in mind, we argue that economics may actually play a heavier/different role in different spaces (layers). For example, the decentralization behind blockchain networks initially motivated a fully decentralized Resource Execution Space. But latency and computing performance have constantly driven those users to deploy a Participation Space (e.g., miners/stakers in Ethereum) in highly performant Resource Execution Spaces (e.g., hyperscalers). Here, the economics drive the decentralized participation, but network and computing performance ends up concentrating/consolidating deployments.

There is more behind speaking about “spaces” here rather than “layers” than this already long post can reasonably address. Suffice to draw parallels: they separate the problem space into different parts, and explicitly discuss how these parts interact – but use the same view of centralization (although the terminology mixes here in this quote) on all layers. Crucially, this shared view with layer separation helps highlight the mutual influence.

#Conclusion

A fair amount of the discussion on centralization (by whichever term) is about aligning mental models more than anything. Where the previous parts of this series have concentrated mostly on technical “layers”, they teased apart this modelling issue. This part, on the other hand, extended this more technically focused discussion to the social layers, in order to highlight unresolved issues in the debates folk are having.

Although it remains to be widely supported, it does appear to be the case that the exercise conducted in this mini-series so far actually yields a model that can be applies across all ten layers. To summarize:

• Centrality is the measure of how “central” a node is to a network.
• “Centralization” as used in various discussions is most closely aligned with the notion of “betweenness centrality” in network theory.
• The centrality of a node changes depending on:
  • The scope of the questions being explored, i.e. whether local influence outweighs global influence or vice versa.
  • The layer one examines.
• Layering is necessary to untangle the above, as layering reveals that a node can be central and not central (albeit at different layers).
  • It also helps show where influence is exerted from one layer to another.
• Some layers to consider are akin to layers 8-10 in that they describe the human networks of power and influence that shape the technical layers.
• “Betweenness” works equally well for those layers as it does for technical ones.
• Therefore, the discussion ought to be less focused on finding new terminology, and more on acknowledging the above.

Now, there are plenty of reasons why the ISO/OSI layer is not universally loved. In some senses, it is very imprecise. It is not my proposal that we should force adoption of this model by adding more meaningful layers to it.

But I do think the above layers 8 to 10 make sense, in much the same way as Guzman et al.’s “spaces” make sense precisely because they avoid having to discuss if the ISO/OSI model is flawed.

In fact, McFadden proposes a different taxonomy in section 3, which the following sections lay out convincingly in detail:

• economic and power centralization
• traffic and infrastructure centralization
• architectural centralization
• service and application centralization

When one is confronted with at least three different approaches to “layering” the problem space, each of which has merit, there is often a reason for that: the model is too simple.

In this case, it probably means that “layers” are the wrong abstraction.

What these three approaches share is that they’re effectively laying out a one-dimensional space of separation, intentionally or not. “Layering” is strict in this approach, “spaces” acknowledge that this strictness is difficult to model well, and “taxonomy” tries to enforce as few constraints as possible, but does so at the cost of the clarity that strict separation provides.

In other fields, it’s often helpful to move away from trying to map things onto a one-dimensional space in such a case, and map out the axes of a multi-dimensional space instead. Usually it then becomes possible for to project each previous system’s terms onto that newly formed space – and discover some new categories!

That seems like a worthwhile exercise to undertake. But it is also well beyond the scope of this blog post.