Comment by sim7c00

the best thing i think i read somewhere is to think of modern cpus as a bunch of computers on a network. each task is transmitted in messages over the network potentially, executed and its result sent back in a packet. despite this not being fully accurate it does give an idea that even in a set of 100 operations, due to caches, latencies and also kernel/hardware stuff happening during these 100 operations however much you try to isolate them, the first operation submitted might see its result sent back last.

this will help to determine or realize better when and where synchronization is needed even though you might not expect it. (or remoddeling of the code).

its extremely hard to writr concurrent code well. i think i've hardly ever managed lockless code that wasnt super bugy (yet! :D). it takes a lot of focus and patience, and tons details about the platform(s) it might run on.