4 ms·

I'm doing a project using Python on AppEngine. AppEngine forces you to use a bunch of single threads doing what is really message passing, but implemented in Py

by pgroves 15y ago

I'm doing a project using Python on AppEngine. AppEngine forces you to use a bunch of single threads doing what is really message passing, but implemented in Python so it's slower than Ocaml. I think this is going to turn out to be a killer feature of Ocaml once people get used to the fact that using 8 cores at once isn't that great of a goal when on-demand computing platforms let you scale much much larger. Pretty soon no one will be all that concerned about how much happens on any one machine. It will be how much gets done per process, and how much gets done per 'cloud'. AppEngine already does this. And it already is a design feature. The GC is not single threaded b/c they're too dumb to figure out how to fix it (not that you were saying that, but people tend to jump to that conclusion). It's because they want the fastest possible single threaded performance, which means not polluting the GC with the kinds of locking mechanisms necessary to support multi-threaded operation.

cageface 15y ago
I think this is going to turn out to be a killer feature of Ocaml once people get used to the fact that using 8 cores at once isn't that great of a goal when on-demand computing platforms let you scale much much larger. I think this point isn't made often enough. There certainly are times when scaling within a single machine is useful but, as is more often the case, if you have to scale beyond a single physical instance you might as well plan for process-oriented parallelism up front.
- oconnor0 15y ago
  Certainly, but there are circumstances - say high-performance math & data heavy stuff - that benefits greatly from a shared memory implementation. And the ease of use of multi-process architectures needs to be improved as well.
  - pnathan 15y ago
    The common case of parallelism is usually fine with process/message passing work IMO. For the other 20% of the time... threads are awesome. :)
  - avsm2 15y ago
    Note that OCaml is just fine for heavy math and data stuff: the garbage collector has a global lock, but threads work just fine for CPU intensive activities. Just don't generate much garbage (which is important for any data-heavy processing), and all is good.

by pgroves 15y ago

cageface 15y ago

I think this is going to turn out to be a killer feature of Ocaml once people get used to the fact that using 8 cores at once isn't that great of a goal when on-demand computing platforms let you scale much much larger. I think this point isn't made often enough. There certainly are times when scaling within a single machine is useful but, as is more often the case, if you have to scale beyond a single physical instance you might as well plan for process-oriented parallelism up front.

oconnor0 15y ago
Certainly, but there are circumstances - say high-performance math & data heavy stuff - that benefits greatly from a shared memory implementation. And the ease of use of multi-process architectures needs to be improved as well.
- pnathan 15y ago
  The common case of parallelism is usually fine with process/message passing work IMO. For the other 20% of the time... threads are awesome. :)
- avsm2 15y ago
  Note that OCaml is just fine for heavy math and data stuff: the garbage collector has a global lock, but threads work just fine for CPU intensive activities. Just don't generate much garbage (which is important for any data-heavy processing), and all is good.