Re: [arch-d] centralization

[Lars-Johan Liman <[e-mail]>]

Yaakov,

Interesting! Thanks!

I follow much of your argumentation, but I kind of came to a soft halt
at the end, so, for the purpose of discussion:

> which means that one large network is worth much more than 2
> half-sized networks.

I think this statement should be qualified a bit, because I think one
parameter is how "worth" is measured. I would argue (and I admit this is
from gut feeling more than from facts or knowledge, so I will not take
offence from being contradicted) that a non-functional large network is
worth less than 2 funcional half-sized networks, at least for some
purposes.

My gut feeling tells me that the cost of keeping a centralised system
consistent rises with its size, and it's probably not linear. Isn't
there a point where the cost of keeping it consistent and reliable
exceeds the extra value added by increasing its size by delta-x?

Or, where the risks of systemic failure overtakes the added value.

Granted: obviously these points are moving targets, being pushed by
development, but still?

The "worth" can be measured not only in short time capital gain, but
also in risks, no?

> Over time this pressure can't be resisted.

Agreed, if measured only in short time capital gain, but maybe it can be
(at least for longer time) if political backpressure (with less focus on
capital gain) is applied?

				Best regards,
				  /Liman

#----------------------------------------------------------------------
# Lars-Johan Liman, M.Sc.               !  E-mail: [e-mail]
# Senior Systems Specialist             !  Tel: +46 8 - 562 860 12
# Netnod Internet Exchange, Stockholm   !  http://www.netnod.se/
#----------------------------------------------------------------------

>> Market consolidation and centralization can be good ...
>> But there are also some serious drawbacks.

[e-mail] 2021-03-11 17:57 [+0000]:
> Yes. There are often trade-offs and these trade-offs can be analyzed.

> In response to one remark I pushed into the chat window a remark about the CAP theorem.
> This theorem is commonly used when analyzing distributed systems,
> and is directly applicable to networks which are distributed systems
> for delivery of packets.

> Basically the CAP theorem says you can't have all three of
> Consistency, Availability, and Partition (fault) Tolerance.
> IETF people are used to distributed routing protocols that emphasize A
> and P at the expense of C
> (which is why we need a TTL field in the packet header),
> while computation academics who expect consistency gave us centralized SDN.
> For databases the two extremes are called ACID and BASE.

> To enable more nuanced trade-offs networks use both control
> (distributed) and management (centralized) planes,
> for example centralized management for policy and distributed control for rerouting.
> (So, what SDN does is not to introduce a control plane, but rather to
> replace the control plane by a management plane.)

> However, when referring to multi-user networked applications (not the
> communications service)
> there is strong pressure towards centralization.
> Why do we have basically one Facebook, one Amazon, etc?
> Because the value of a network application is super linear (Metcalfe's
> law puts is a N^2, Reid's law puts it at 2^N)
> which means that one large network is worth much more than 2 half-sized networks.
> Over time this pressure can't be resisted.

> Y(J)S