Concurrency should always be studied from higher-level logical program/algorithm execution properties with a focus on both shared-memory and message-passing architectures. Only after the above should one even look at lower-level hardware memory consistency model, cache coherence model, atomic instructions, memory barriers etc. For example, learning to use a logical shared-memory "Mutex" is far easier than the fact that it is internally implemented using atomic instructions, memory barriers etc.

Some relevant details in a previous comment of mine - https://news.ycombinator.com/item?id=42422867

See also this book - https://news.ycombinator.com/item?id=42472334

> Concurrency should always be studied

I disagree. If we want people to understand, then we should teach computer systems bottom-up and thus understand the bottom-level behaviors before being surprised that our bad (i.e. wrong) abstractions are leaking later. If you want to train programmers to accomplish parallel programming tasks via plugging together nice abstractions, that's all well and good, but that's a completely different kind of instruction than trying to teach them how things work. If we only do the latter then we'll never have experts again.

In my previous comment i linked to an earlier comment of mine which already pointed it out, so this is not revelatory. W.r.t. concurrency however, the logical approach should precede lower-level implementation approach since there is lots more complexity involved in the latter. Mutexes/Semaphores/Condition Variables etc. are not bad/wrong abstractions but necessary correct ones. They are fundamental to understanding/managing concurrency itself.

The issue is that logically correct algorithms can indeed be made, but actually making them work in real conditions is its own problem. It's not necessarily wrong to explore more exotic algorithms, but a focus on algorithms that people should actually use in their given situations seems higher priority. We should be working on bringing useful concurrency to our programs right now.

And while I will agree that mutexes and the like are useful in their own right, they are not the most fundamental for understanding concurrency.