- 130comments
- 3comments
- 10comments
- 192comments
- 41comments
- 172comments
- 2comments
- 25comments
- 59comments
- 409comments
- —discuss
- 89comments
- 40comments
- 53comments
- 158comments
- 637comments
- 28comments
- 3comments
- 264comments
- 99comments
- 177comments
- 86comments
- 10comments
- 240comments
- 55comments
- 163comments
- 168comments
- —discuss
- 25comments
- 31comments
Although the authors have designed hardware for this, I can imagine it being used by a language's VM (e.g. the JVM), making use of cache size-aware Zippads.
In the article they mention that most programs are not dominated by arrays but objects. An other cure for that is to use arrays of primitives a lot more:
https://www.youtube.com/watch?v=yy8jQgmhbAU
Another great talk on data-oriented design, given by Mike Acton at cppcon 2014 https://www.youtube.com/watch?v=rX0ItVEVjHc
I once wrote an interpreter that performed some operation on interpreting a data structure. Because it seems that the interpreter was spending most of its time on interpreting the data structure, I thought it would be a good idea to 'compile' the data structure in code. That resulted in a lot of code, which took a long time to compile, but I was hoping that the compiled code would be much faster. But it turned out to be almost twice as slow. Then if dawned to me: The interpreter and the data structure fitted in the cache, while the compiled code did not.
I bet that in some cases using a dedicated memory allocator that takes care that objects that belong together are stored together, can result in execution improvements. If you using the default memory allocator, it could happen you get pieces of memory far from each other, especially if other threads are also allocating memory or because temporary objects (think string manipulations) are created during the construction of the data structure.
Adapt a perfectly working machine to a mental crutch instead of learning data oriented design?
If this means running millions of already written programs more efficiently - yes!
The question is whether those gates are really best used for this feature. Also, this feature might increase latency and make tens of millions programs run slower.
It probably can be done generic enough in a cpu, reordering data layouts on the fly using some access and temporal locality tracing like done for caches. Slight memory cost.
Security implications are important though, with programmer nor kernel no longer controlling memory layout.
Initial stages of this are already seen in the various ways of memory interleaving for multiple banks, cores and cpus.
The truly fun part would be speculative as opposed to tracing reordering...
Compression could potentially help bandwidth, but it isn't going to do anything for memory latency, which is the actual problem. That being said, if compression is part of a CPU cache it could indirectly help, but then I'm not sure how that is going to be done on 'objects' (whatever their definition is).