parallelFuture

[dsimcha]

== Quote from Charles Hixson (charleshixsn at earthlink.net)'s article
> dsimcha wrote:
> > I've created an alpha release of parallelFuture, a high-level parallelization
> > library for D2.  Right now, it has a task pool, futures, parallel foreach, and
> > parallel map.
> >
> > Here's the (IMHO) coolest example:
> >
> > auto pool = new ThreadPool();
> >
> > // Assuming we have a function isPrime(), print all
> > // prime numbers from 0 to uint.max, testing for primeness
> > // in parallel.
> > auto myRange = iota(0, uint.max);
> > foreach(num; pool.parallel(myRange)) {
> >     if(isPrime(num)) {
> >         synchronized writeln(num);
> >     }
> > }
> >
> > The interface is there and it seems to work, although it has not been
> > extensively stress tested yet.  Some of the implementation details could
> > admittedly use some cleaning up, and I would appreciate help from some
> > threading gurus on improving my queue (right now it's a naive synchronized
> > singly-linked list) and getting condition mutexes to work properly.  (Right
> > now, I'm using atomic polling followed by sleeping for 1 millisecond in a lot
> > of places.  It's a kludge, but it seems to work reasonably well in practice.)
> >
> > The code is at:
> >
> > http://dsource.org/projects/scrapple/browser/trunk/parallelFuture/parallelFuture.d
> >
> > The docs are at:
> >
> > http://cis.jhu.edu/~dsimcha/parallelFuture.html
> If you can easily do it, it would be nice to be able to have the threads
> able to communicate with each other.  Something along the lines of both
> broadcast messages and 1:1 exchanges.  A very rough sketch of a possible
>   use:
> Task[] tasks  =  Task.who_is_waiting;
> foreach (auto task; tasks)
> {  if (task.process(something) )
>     {  task.markDone(true);   }
> }
> I see "something" as being an Object that would need to implement an
> interface to carry some identifying information, so when a task received
> it it could determine what to do with it (with "reject processing" being
> an option).  I think the rest is pretty clear.
> OTOH, this comment was just to demonstrate the kind of communication
> between tasks that I'm talking about.  Static methods for broadcast
> communication and instance methods for 1:1 communication.  With
> safeties, so when a task completes before it gets your message, you
> aren't left talking to a null pointer.  Cleanup would need to be managed
> by the original thread that created the tasks.  It would need to be able
> to send a broadcast "now closing task xxx signal" to all threads.
> Threads maintaining a list of tasks would need to scan through them and
> remove any references to that task.
> Maybe this is getting to complicated, but it would certainly be useful.
>   In the past interthread communication is the main problem that's kept
> me from using them.

I'll think about this and see if I can work something like this in, but I need
real-world use cases.  I do bioinformatics work, which basically means large-scale
data mining, embarrassingly parallel problems and very CPU-bound work.  The use
cases I had in mind were mostly the "use every core I have to do something
embarrassingly parallel" kind.  The goal was to make these use cases as dead
simple as possible.

Whatever use cases you have in mind, I'm apparently not familiar with them.  If
I'm to improve this lib to handle use cases other than the pure
throughput-oriented parallelization of embarrassingly parallel tasks that I had in
mind, I need to understand use cases from other fields.