Re: [Caml-list] Distributed computing libraries

[Francois Berenger <berenger@riken.jp>]

On 06/07/2012 05:52 AM, Thomas Braibant wrote:
>> These are very different pieces of software, because they tackle problems at
>> various abstraction levels. ocmc is the most level-level here, as it "only"
>> tries to improve the runtime so that threads can be run in parallel on
>> multiple cores. That's it, there is no additional abstraction on top of the
>> standard threading API - no distribution, no computing.
>>
>> JoCaml uses the normal multi-threading in the runtime, but integrates it
>> differently into the language. So, it adds abstraction, but you are still
>> limited to a single core.
>>
>> So far I know, all the other libraries base on multi-processing to run
>> programs on multiple cores. Plasma is the only one with true distribution
>> capabiltiies beyond a single computer, but the price is that you must use
>> the map/reduce scheme, whereas functory or parmap leave you more freedom.
>> However, all multi-processing approaches share the property that the data
>> flow is limited by process boundaries (unless you go really low-level and
>> also take Netmulticore into consideration (part of Ocamlnet), which uses
>> shared memory to overcome these limitations).
>>
>> I don't know what you are exactly looking for. Knowing the problem it would
>> be easier to recommend something.
>
> Thanks a lot for this answer.
>
> As I said in another email, I do not have a problem to solve right
> now. But, I was tickled by the Parmap announce,

Don't hesitate, jump on it, that's really a nice technology.
The "minimal disruption" concept is quite interesting.
For example, if your List.iter is changed to a Parmap.pariter,
the parallelization of this portion of the code is done.

It's quite comfortable to develop and debug in single core mode 
(List.iter) and switch to the // version only once you're happy with
the sequential one.

Regards,
F.

 > and I wondered how it
> fared with the other related systems I knew about. (Yet, I reckon that
> I should have made the question more precise, with a detailed lists of
> criteria, akin to what Oliver suggested.)
>
> In my mind, asking on the list was the best way to make a good summary
> of the current situation on this front. (But I am afraid I do not know
> the best place to make public this kind of summaries.)
>