Comment by einhverfr | Hacker News Reader

Comment by einhverfr | Hacker News Reader

Parent Full thread

One of the things I have often noted is how theories which don't match reality very well often trump reality in terms of explanations. For example, TCP/IP is a 4-layer network stack, while the OSI model is 7 layers. These models are fundamentally different and understanding the differences is often key to understanding design decisions for OSI protocols ported to TCP/IP (H.323 being a good example).

But people talk as if TCP/IP follows the OSI model. It doesn't. Often understanding that it doesn't and why it doesn't is a good first step to understanding why it works the way it does. So that is top of my list.

THere are other issues. For example, when another techie is speaking gobbldiegook, it is way too easy to project meaning onto what is said so that misunderstandings develop. For example, I have written frameworks for building object frameworks. These are not object factory factories (but rather services to offer object frameworks, for example generalized service locator API's for locating stored procedures, allowing a looser coupling than most people have). However, when someone says "its an object factory factory" very often the simplest response is "well, sort of, if you are the factory worker!"

Don't be baffled. When was the last time you went to the movies to see the ordinary life of that normal person called Tony Stark. Buzz is what we desire. As for the TCP/IP and OSI differences, you are being slightly iffy. While they are definitely designed differently, TCP/IP only coalesces OSI layers to make programming simple for the task at hand.

> While they are definitely designed differently, TCP/IP only coalesces OSI layers to make programming simple for the task at hand.

But it doesn't except if you use the cliff notes version of the OSI model. That's my point. Otherwise, H.323 would look like SIP, which it doesn't. Protocols designed for an OSI model look fundamentally different than protocols designed for a TCP/IP model.

Let's look at H.323 since that is the example I am using. The protocol is a beast to work with over TCP/IP because it was not designed for that. It was instead designed for a unified OSI network where virtual circuits could be requested to handle voice and video information. Those don't exist in TCP/IP which has no concept of a virtual circuit and so over TCP/IP, you use a new TCP connection everywhere you'd use a virtual circuit because that is the closest equivalent. It isn't really an equivalent however which makes the protocol very, very difficult to deal with in terms of firewalls, etc. Compare with SIP which just passes everything across a single TCP connection hitting a well known port.

The difference here is that the basic assumptions of what the network can do are different. TCP/IP assumes at its basis that you have a simple, dumb packet switched network that really isn't capable of doing anything else. OSI assumes you have an intelligent network capable of routing packet-switched and circuit-switched information together. They are totally different design criteria and at best you can say TCP/IP is a very narrow subset of what OSI is designed to do.

What this means, effectively, is that you either have to strip down the OSI network in theory to match TCP/IP (which is useless) or you end up with something that doesn't really match. Either way there is no way of understanding H.323 without understanding the OSI model to a level in which the fundamental differences with TCP/IP matter and that requires knowing the basics not only of packet-switched networks but also circuit-switched networks too.

How about we simply agree that OSI is a monstrous "designed by 99 committees" theory and accept that it is not used in practice ? Actual networking stacks do not follow it. Hell they don't even follow TCP/IP's layers, but at least there they make an effort to make it look like they do (e.g. in networking hardware, or for that matter the linux kernel, there is no actual "layer 3 routing" step that is distinct from the "layer 2 switching" step).

OSI is an ancient, dumb, wrong, 10000000000000 feet architect view of networking. Do not learn anything about it, you'll do yourself a favor. Ethernet is NOT OSI layer2, IP is NOT layer3, and especially do not think TCP is OSI layer 4.

> How about we simply agree that OSI is a monstrous "designed by 99 committees" theory and accept that it is not used in practice ?

Partly agreed. It's worth noting however that the PKI associated with SSL (X.509) is an OSI thing and so understanding OSI concepts is still rather important. This is also why X.509 certificates IMO feel a bit foreign in terms of structure and they tie in closely to LDAP concepts (a stripped down version of OSI X.500) rather than more internet-y directory concepts like DNS where everything is a text string and human readable.

> Hell they don't even follow TCP/IP's layers, but at least there they make an effort to make it look like they do

My point was that TCP/IP and OSI are just fundamentally different. They are culturally different. They have different goals. You can understand TCP/IP better on its own than through OSI. You can't understand OSI by using TCP/IP as a sole reference.

> "OSI is an ancient, dumb, wrong, 10000000000000 feet architect view of networking."

I think it is more correct to understand OSI in context, namely that it was a network model designed to support the OSI protocol stack. This stack was intended to fully merge circuit and packet-switched networks (effectively merging phone switches and the internet). This means you have, effectively, an intelligent network with possibly intelligent terminals, which are capable of doing both circuit- and packet-switched networking on top. OSI protocols ported to TCP/IP like H.323 make perfect sense if you understand that they are assuming they can open phone calls (but in the process of porting them to TCP/IP, they have to use one TCP/IP connection for each phone circuit they would otherwise open). This is furthermore why the network perimeter controls designed for H.323 involve what amount to phone switchy things instead of things like firewalls (and why a simple packet filter won't work).

In terms of the value of not learning it, I think the best approach honestly is to learn TCP/IP without the OSI model first, and then learn the OSI model and what the OSI stack was designed to do after. This gives you whole layers of insight, not into TCP/IP but into OSI protocols ported to TCP/IP.

Maslow's hierarchy of needs is similar. It's not true in an absolute sense, but it is a convenient scaffolding for understandable explanations at a knowledge appropriate level, and that has a value beyond simple truth.

Agreed.

Interestingly Roberto Assagioli makes a case that Maslow's hierarchy of human needs is useful beyond that, not so much in terms of the heirarchy itself so much as the idea that peak experiences (including mystical experiences) are things which are easiest to achieve once certain psychological things are dealt with.

Assagioli is interesting reading. It is as if he throws Freud, Jung, and Dante into a pot, seasons with a touch of Hinduism and Arthurian legend and pulls out out a psychological theory which surprisingly makes sense.

Reply on news.ycombinator.com