[ih] Speaking of layering and gateways

[Jack Haverty]

On 4/16/24 02:21, Johan Helsingius via Internet-history wrote:
> On 16/04/2024 07:46, Brian E Carpenter via Internet-history wrote:
>> Especially - bits can be duplicated and retransmitted at low cost;
>> furniture can't.
>
> In my last home move, a lot of the furniture ended up in bits...
>
>     Julf
>

I agree that analogies have a limit.   But even "stuff", like furniture, 
can be and is retransmitted.   Occasionally when I receive a package 
that has been mauled in transit, I report it to the sender, who usually 
promptly sends me another one.  No doubt insurance companies might be 
involved too, who investigate who to blame and allocate costs 
accordingly.   My "packet drop rate" for packages is very small but 
non-zero, and the "error detection" is done by me at the endpoint.   If 
the box arrives with a hole in it, extra inspection is warranted.   I 
suspect also that if a "drop rate" climbs, the sender makes changes to 
its packaging and shipping processes to get the losses back down to an 
acceptable level.

In the Internet universe, it was useful to be able to monitor such 
packet (errr, datagram) behavior.   SNMP mechanisms were defined long 
ago to glean data from TCP endpoints, where statistics about drops, 
duplicates, retransmissions, checksum failures, and such were accessible.

When I was involved, 30 years ago, in operating a corporate internet 
(intranet), we actually gathered such data especially when trying to 
figure out incidents reported by users' complaints such as "the net is 
really slow today".   I recall instances where we used that data to 
isolate error-prone, but still functional, circuits. Transpacific 
circuits were especially problematic, at least 30 years ago.   In other 
cases, we sometimes found bugs in a vendor's TCP implementation.   You 
can't trust that an implementation actually behaves as the RFC 
specifications dictate.  TCP's robustness is actually a detriment in 
such situations.   TCP's mechanisms keep data flowing, but also hide 
problems until they become very severe.

I also recall an Internet meeting, back in the 80s, where we debated the 
question "What should be the "normal" drop rate for IP datagrams"?   
Someone eventually shouted out "One percent", and the group quickly 
reached a consensus.   No analysis, no OR, no equations or models were 
involved.  It just seemed right to the group.

I'm glad to see that there has been some work to apply science such as 
Operations Research to the Internet world.  I remember from my short 
exposure to OR that it was often used to improve decision-making.  
Decisions might where to add a new airline route, or whether to expand 
an existing manufacturing plant capacity or build a new plant somewhere, 
and where to build it.  In operating a distributed system, there are 
many such decisions to be made - such as what is a "normal" drop rate 
for IP datagrams.   Or perhaps "How long can a datagram be in transit 
before the Internet is considered broken for that user?"

During my stint as a network operator, we never could find the "How to 
Operate Your Internet" manual.  Decisions were made largely by instinct 
and intuition rather than scientific analysis.

Some of our decisions were made using economic data, which was the only 
data readily available.  For example, there was considerable demand for 
traffic between our sites in Europe.  In one case, two offices could 
actually almost see each other across a river.   But they were in 
different countries and cross-boundary circuits were extremely 
expensive.   At least at the time, it was much cheaper to run a circuit 
from each country (in Europe) to New York City.   So all traffic, almost 
line-of-sight "across the river" in Europe, actually crossed the 
Atlantic, twice.  That greatly increased latency, but in those days we 
weren't doing things like Zoom so it wasn't an issue.

How are such decisions reached today in the Internet?  What's the 
process for deciding where to put an IXP?  Or where to run a new fiber?  
Or where to put your equipment to build a cloud?   Or how to select 
which ISP to connect your house or office to the Internet?

Perhaps someone has insight into how such decisions are performed today 
in the Internet, how scientific methods such as OR are used, and how 
such decision processes have changed over the 50 years of Internet 
evolution.  Then compare to how its done in other environments such as 
transportation of goods.  This might make a good topic for some 
university thesis.

Jack Haverty

-------------- next part --------------
A non-text attachment was scrubbed...
Name: OpenPGP_signature.asc
Type: application/pgp-signature
Size: 665 bytes
Desc: OpenPGP digital signature
URL: <http://elists.isoc.org/pipermail/internet-history/attachments/20240416/43352915/attachment.asc>