www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - Next computer architectures & D

reply bearophile <bearophileHUGS lycos.com> writes:
If you are *physicist* and you look at the future of computer architectures,
you can think that it's inevitable that sooner or later multi-core computers
will adopt an architecture where memory is not shared, each core has its own
memory (plus a bit of memory for message caching), and sends messages to other
cores. For such hypothetical computer architecture an actor-based programming
language is quite fitting. So it's good for D to have something like actors.

But several practical considerations make such model appear unrealistic. So I
think a structure closer to the reality will be the same with units that don't
share a single memory and send messages to each other. But the units themselves
will be composed by several cores that share a single memory (plus caches), and
each core will have advanced SIMD instructions (see AVX instructions that
perform a kind of vectorized 'if' too).

(Beside such structure there can be something like a GPU, with many more tiny
cores, see the recent NVIDIA Tesla.)

So a programming language that wants to be efficient needs to use well such
two-level architecture (the Chapel language seem able to adapt to that, it has
domains for its memory, etc.). So maybe an actor-like-based concurrency design
is not enough. D has array ops that once improved and well implemented can be
use SIMD instructions in a good enough way.

Bye,
bearophile
Jul 27 2010
parent reply BCS <none anon.com> writes:
Hello bearophile,

 So I think a structure closer to the reality will be the
 same with units that don't share a single memory and send messages to
 each other. But the units themselves will be composed by several cores
 that share a single memory (plus caches), and each core will have
 advanced SIMD instructions (see AVX instructions that perform a kind
 of vectorized 'if' too).

I'm no expert but that sounds a lot like what I know about current generation super-computers. -- ... <IXOYE><
Jul 27 2010
parent bearophile <bearophileHUGS lycos.com> writes:
BCS:

 I'm no expert but that sounds a lot like what I know about current generation 
 super-computers.

But I think that architecture will be seen inside PCs in ten years or less :-) It's being years that Intel is able to create a CPUs with 30-80 cores, but it's kind of impossible to keep efficient a single coherently shared memory with that many cores. You need to do as GPUs do. But I think such flat architecture is not so good with a large amount of legacy code and it's not as efficient as possible with tons of partially serial code. So I think you will see the more complex hybrid I have explained. And I think D is not ready yet for such hybrid architecture yet. Bye, bearophile
Jul 28 2010