Hacker News Reader

Comment by Sharlin | Hacker News Reader

Not just allocations, but critically – especially on modern machines – cache coherency. Indirections are an enemy of performance, and indirections to many different parts of memory in particular. A custom allocator can be a great help though, especially in a compiler where you can pretty much just use a fast slab/bump allocator and not worry about freeing memory until the end.

View on HN

Do you mean cache locality? Coherence is not something ordinary single-threaded code has to be concerned with.

Oops, indeed, brain fart.

We have terms and languages to express computational complexity, algorithms, temporal logic etc.

But what about data logistics and CPU mechanics?

How good are we at expressing things like memory fragmentation, cache locality, throughput and latency of operations etc? All of these things affect how the machine operates in very real ways, to a degree that often trumps the higher level mathematical concepts above.

I only ever see these things explained in high-level, heuristic terms, observed as a black box or spelled out literally as code.

Is there a language based on how the CPU and memory work that I'm not aware of?

I do not think that such a language has been developed. The complexity of how a CPU, with several levels of caching, RAM, harddrives (with their own caching), and network storage (with their own caching) is simply mind boggling.

For that reason, I think that this route is not often taken. I guess that whenever modeling of execution is performed it is done to analyse the behaviour of an existing implementation in order to find ways to improve it.

If there is someone who has made an attempt to do something in this direction, I think it is Donald E. Knuth in his 'The Art of Computer Programming'. I think, for example about section 5.4 'External sorting' in Volume 3 'Sorting and Searching' [1].

[1] https://seriouscomputerist.atariverse.com/media/pdf/book/Art...

Reply on news.ycombinator.com