Re: [Caml-list] Async and lwt

[Yaron Minsky <yminsky@janestreet.com>]

The conceptual differences are pretty thin; there are some error
handling differences, but they're not huge.

As for the shared heap question: Async greatly improves the
programmers ability to reason about concurrent code by providing clear
bounds on when interleaving can happen.  In particular, a given job
scheduled to run as part of Async can never be interrupted by another
async job. That means interleavings can effectively only happen where
you use a bind operator to string two closures together, and so one is
rarely exposed to race conditions. That means that Async program
typically don't need to think much about locking, and race conditions
stop being your primary source of bugs.

Using Async in the context of shared heap parallelism gives up these
guarantees, and puts you back in mutex/sempahore/condition-variable
hell.  My guess is that we'll end up taking advantage of the multicore
GC by running multiple Async schedulers, one per domain (e.g., OS
thread), and having carefully written primitives for sharing
data-structures and efficiently sending immutable messages between
these domains.  But I think just freely scheduling Async jobs across
multiple physical threads seems like a disaster from the point of view
of producing comprehensible, reliable code.

y

On Thu, Jun 30, 2016 at 8:41 PM, Hendrik Boom <hendrik@topoi.pooq.com> wrote:
> On Thu, Jun 30, 2016 at 08:13:58PM -0400, Yaron Minsky wrote:
>> I'm not at all sure that the decoupling is possible or wise for Async.
>> My intuition is that this is too complex of a problem with too much
>> need for careful optimization to be able to have a simple, shared
>> generic data structure for this.
>>
>> The solution that seems most plausible to me is to settle on one
>> implementation, and port the API of one library to run on top of the
>> other. There was indeed an experiment in this direction that was done
>> by Jeremie Dimino:
>>
>> https://github.com/janestreet/lwt-async
>>
>> That said, until we resolve the binary size issues with Core and
>> therefore Async, I doubt that this solution would be appealing to the
>> full community of lwt users.
>
> What are the conceptual differences between Async and lwt?  Does either
> of them manage to take advantage of a shared-heap multicore system?
>
> -- hendrik
>
> --
> Caml-list mailing list.  Subscription management and archives:
> https://sympa.inria.fr/sympa/arc/caml-list
> Beginner's list: http://groups.yahoo.com/group/ocaml_beginners
> Bug reports: http://caml.inria.fr/bin/caml-bugs