This question is motivated by the discussion in this GitHub issue.
I asked Edge Copilot the same but I don't fully trust LLMs. Here is an excerpt from that chat
Aeron: Provides a high-performance, low-latency transport layer. It can operate over TCP or UDP, optimizing data transmission for speed and reliability. It’s designed to be more efficient than using TCP/IP or UDP directly, especially in scenarios requiring high throughput and low latency.
ZeroMQ: Provides a messaging layer that abstracts the complexities of socket programming over TCP/IP. It focuses on messaging patterns and ease of use, making it simpler to build distributed applications.
So, while ZeroMQ abstracts the transport layer to simplify messaging, Aeron optimizes the transport layer itself to improve performance. You can think of Aeron as a specialized tool to make data transmission faster and more reliable, which can be used in conjunction with messaging systems like ZeroMQ.
So somehow Aeron both uses TCP and optimizes TCP? Has Copilot mislead me?
Q1 :
" What is the distinction between transport and messaging? "
Well, as 've spent more than a decade programming & consulting system designs with ZeroMQ, since about v2.0.11, many people I tried help with this same question started to distinguish the concept of a messaging & signalling layer as the "product" of all the hidden work below. These top-level abstractions "move-data-payloads" in some application level context ( some signal self-healing, auth-reauth claims, some detect errors to help self-healing of the application-level logic - some others simply pass data or interact in some data-driven distributed application work-flow logic ).
On the very opposite end, almost on the bare metal, be it for wires or wireless, we have the means of transport ( for ages I promote calling this side abstractions a set of available Transport Classes rather than "protocols", as ZeroMQ is capable of working, at the same time, from the same PUB-( publishing role archetype )-AccessNode, using many different links with different TransportClasses ~ at the same time ~
{ inproc: | ipc: | tipc: | tcp: | vmci: | pgm: | epgm: | norm: | ... }
So the transport ought be felt as a lowest-level abstraction for means actually "mediating" use of technical channels, built among distributed AccessNode-s ( a PUB-role-AccessNode, PAIR-role-AccessNode, ... ) to somehow communicate ( ISO-OSI-Layers down to ..., -3 Network, -2 LLC, -1 PHY ) over actual signal-delivery medium with it's counterpart on the other end of the "wire".
Q2 :
" (...) both uses TCP and optimizes TCP? "
ROFL
It is fair not to expect (any) Large-Language-Model to do anything else, than a programmed mere Phrase-Transformation ( given language-grammars + source texts the Phrase-Transformation code does nothing else ).
Phrase-transformation, no matter how advertised, just meets the grammar ( better to say a subset of the grammar it "saw" in the source-code or "met" in the corpus ) and mechanically follows the statistics of "appearances" in the source-texts corpus, nothing more.
Zero "understanding" what was actually written in the corpus.
Zero "comprehension" what the texts in corpus actually meant.
Who, reasonably thinking, would ever expect a Finite-State-Automaton ( yes, even though having a large State-space, the LLM is still nothing more than a constructed FSA, underscored F-inite, highlighted A-utomaton ) to indeed eliminate logical nonsense, like the cited "uses TCP and optimizes TCP", when it "read" each part of this assembled nonsense in the source texts? It would require both understanding and comprehension and thinking ( critical thinking ), neither of which the programmed Phrase-Transformers are capable of.
Source texts in the training corpus were collected ( not selected, like when creating any evolving state-of-knowledge preserving titles, like Encyclopaedia Britannica and similar curated corpora ). Merely lowest hanging fruit ( cheapest, ready to machine-reading, yes, internet visible ) texts got collected, while written by different people - marketing, opinion-makers, advertisers, urban storytellers, blogophilles, be them knowledge-focused or "attention"-grabbers + in recent years by more and more media-pretending bots ( re-publishers of some other source texts ). Put together, the Phrase-Transformer takes it all, and only re-formulates what has been already written.
What is an Added value from such "FSA-recycled-phrases"?
"Amount" of FSA-generated texts? Phew...
"Speed" of FSA-generating whatever? Phew...
"Cost" of FSA-generated smog? Phew...
Ol' phrase says "If a thing costs nothing, it has no actual value."
Q3 :
" Has Copilot mislead me? "
No, a Finite-State-Automaton has not.
It was your feeling you might get something ( of some real value to you ) from a mere Phrase-Transforming Finite-State-Automaton.
Do not panic. Human were for ages fascinated by things, that seemed to do something useful "automatically" ( like 18-th century mechanical chess-machines, like 19-th century "almost human" marionettes, like 20-th century "dancing"-robots, like 21-st century "chat-capable"-robots, like tactile robots (in academia demos, failing otherwise) ) or "for free", so this recent hype of still just mere FSA toys ( be it called GPT, co-pilot, assistant, A.I., whatever marketing people invent tomorrow ) are nothing more than a "reincarnation" of the same trick - to make you "believe" it can do something you might wish it to do, but technically not being there (yet).