upvote
> keeping that illusion consistent and performant requires deep cache hierarchies and coherence protocols that cost area and power, and scale poorly.

Cache hierarchies also exist because memory latency and capacity tend to be inversely proportional to each other, and directly proportional to cost. The underlying process and physics of chip design works against you on this front.

I'd love to have 1 GB of L1 cache, but that is simply not viable because your cache would be as big as your core, way more expensive, and by the sheer size, it would become a NUMA cache, which defeats the point of a cache and the "illusion" of UMA.

reply
Yes, there’s a notable IEEE article from like 15 years ago that riffs on these same points—the gap between the abstraction presented by the typical ISA and the actual microarchitecture is too great to ignore and we have blindingly complex and opaque layers of compiler optimizations and OOO pipelines that exist solely to maintain the fiction for programmers.

I think the greatest benefit of rethinking CPU architecture at the level of the ISA will be to make performance more transparent to developers without the layers of black magic.

reply
> Computers don't have global memories anymore, they pretend they do.

Citation needed. After 10 minutes of searching I can't find anything to support this statement.

reply
> It's hard because you made a fatal assumption: There is only one global memory. How exactly is anyone supposed to reason about such an omnipotent and all powerful concept, where everything can potentially affect anything?

Because parallelism is not that needed, there’s always contention starting from physical constraint. A user can only have one body, a printer can only output one piece of paper at a time,… There are natural areas for parallelism and it’s been already implemented. Memory as a bottleneck has not been a things for decades.

Breaking down things in terms of tasks is not hard. The hard part is discovering that those parts need to communicate. And sometimes they want access to same limited resources. We already have things like array programming that have a parallel view of computation. But no one is rushing to get things done with them.

reply
I'm with you, but I don't see a clear path from "here" (illusion of the same memory) to "there" (each processor has its own memory, I guess). It's really hard to reason about latency, synchronization and communication costs in the abstract.

In a way I see the various cloud computing services as the organically grown (i.e. haphazard and clunky) answer to this problem.

reply
One of the reasons it is so hard is because the discipline of thinking about those things isn’t baked into the language like it should be. You can’t look at a piece of code in a typical high-level language and reason about synchronization overhead because those semantics aren’t expressible in the code. We need languages that support expression of effects and coeffects, which would enable that kind of high-level reasoning about tradeoffs.
reply
Do you have any ideas on what that would look like? The best I can come up with is something like https://seastar.io/ built in to the language.

But even that feels lacking, even though I'm really impressed with Seastar. I love the "bring operations to data" thinking.

reply
I agree, I've always viewed C as a sort of domain specific language for Von Neumann architectures. We only call it a "general purpose language" because the domain was so big, but using C to write multi-core distributed, networked programs to me feels like using CSS to write an operating system.
reply