Hacker News Reader

Comment by rayiner | Hacker News Reader

The "memory is just a giant array of bytes" abstraction hasn't been true ever since DRAM has existed (because DRAM is divided into pages), ever since caches were introduced, and certainly isn't true now that even commodity mulch-processors are NUMA with message-passing between memory nodes.

View on HN

Look if we want to be super anal about shit all memory is slowly discharging capacitors with basically random access times based on how busy the bus circuity is with our shit this week. It turns out that memory is really complicated stuff if you look at it deeply, but the magic of modern computer architecture is that you get to (hopefully) keep your shelf model for as long as you can. If you were to try to model actual memory latency: here's a shortcut: you can't. That's why everyone bullshits it.

Fair point. At this point, it's a very leaky abstraction because not all levels of "random access" (e.g. L1 cache vs. main memory) are created equal.

True, and this is my biggest problem with writing optimized code in C -- it takes a lot of guessing and inspecting the generated assembler and understanding your particular platform to make sure you're ACTUALLY utilizing registers and cache like you intend.

If there were some way of expressing this intent through the language and have the compiler enforce it, that'd be fantastic :)

That said, there's really not a better solution to the problem than C, just pointing out that even C is often far less than ideal in this arena.

i'm actually spending much of my time these days having a go writing HPC codes in haskell.

So far, the generated assembly looks pretty darn awesome, and the performance pretty competitive with alternatives i have access too :)

Reply on news.ycombinator.com