Jump to page: 1 2
Thread overview
parallelFuture
Oct 22, 2009
dsimcha
Oct 22, 2009
Tim Matthews
Oct 22, 2009
dsimcha
Oct 22, 2009
Tim Matthews
Oct 23, 2009
dsimcha
Oct 24, 2009
dsimcha
Oct 24, 2009
dsimcha
Oct 24, 2009
Christopher Wright
Oct 24, 2009
dsimcha
Oct 22, 2009
zsxxsz
Oct 22, 2009
dsimcha
Oct 22, 2009
Charles Hixson
Oct 22, 2009
dsimcha
Oct 22, 2009
Charles Hixson
October 22, 2009
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
October 22, 2009
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

Nice. About the tasks:

In .Net a worker thread is created for each cpu core. Each worker thread has its own local queue of tasks which it initially retrieves from the global queue but if the tasks creates new tasks it adds to its local queue directly for less contention. When it finishes completing a task it takes the next from the back of its local queue to take advantage of cache (like a stack). The tasks at the front of the queue can be stolen from another worker thread if its local queue and the global queue are both empty to maximize cpu usage.

Also the task's wait for complete is only considered complete if all of the tasks it created too are complete (kinda like recursion).

What is the implementation or plans like for this.
October 22, 2009
== Quote from dsimcha (dsimcha@yahoo.com)'s article
> 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

Very good!
October 22, 2009
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


Very nice! Parallelisation for us ordinary folks. :) I tried your isPrime example, which was very cool. I can't wait to try the library in a real application.

I often have a situation where I have a set of grid points, and I do (mostly) separate calculations on each grid point. Your library should allow me to do calculations on several grid points in parallel, with minimal changes to my code.

Is there some particular reason why you have capitalised the F in the file name, but not in the module name?

-Lars
October 22, 2009
== Quote from Lars T. Kyllingstad (public@kyllingen.NOSPAMnet)'s article
> Is there some particular reason why you have capitalised the F in the
> file name, but not in the module name?
> -Lars

This is called the effects of being in hack mode late at night.  I guess the convention is all lower case.
October 22, 2009
== Quote from Tim Matthews (tim.matthews7@gmail.com)'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
> Nice. About the tasks:
> In .Net a worker thread is created for each cpu core. Each worker thread
> has its own local queue of tasks which it initially retrieves from the
> global queue but if the tasks creates new tasks it adds to its local
> queue directly for less contention. When it finishes completing a task
> it takes the next from the back of its local queue to take advantage of
> cache (like a stack). The tasks at the front of the queue can be stolen
> from another worker thread if its local queue and the global queue are
> both empty to maximize cpu usage.
> Also the task's wait for complete is only considered complete if all of
> the tasks it created too are complete (kinda like recursion).
> What is the implementation or plans like for this.

For now, parallelFuture was designed with a single producer, multiple worker model.  Absolutely no attempt was made to allow for tasks running in the task pool to themselves submit jobs to the same task pool, because it would have made things more complicated and I couldn't think of any use cases.  I designed the lib with the types of use cases I encounter in my work in mind.  (mathy, pure throughput oriented computing on large, embarrassingly parallel problems.)  If someone comes up with a compelling use case, though, I'd certainly consider adding such abilities provided they don't interfere with performance or API simplicity for the more common cases.

To make this discussion simple, let's define F1 as a future/task submitted by the main producer thread, and F2 as a task/future submitted by F1.  The queue is (for now) strictly FIFO, except that if you have a pointer to the Task/Future object you can steal a job.  When F1 submits F2 to the queue, F2 goes to the back of the queue like anything else.  This means when F1 waits on F2, it is possible to have a cyclical dependency (F1 waiting on F2, F2 waiting for a worker thread populated by F1).  This is mitigated by work stealing (F1 may just steal F2 and do it in its own thread).

In parallel map and foreach, I should probably document this, but for now it's undefined behavior for the mapping function or parallel foreach loop body to submit jobs to the task pool and wait on them, and in practice will likely result in deadlocks.
October 22, 2009
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.

October 22, 2009
== Quote from Charles Hixson (charleshixsn@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.
October 22, 2009
dsimcha wrote:
> == Quote from Charles Hixson (charleshixsn@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.
It's basically a simplified form of message passing, and useful for general processing in the context of multiple cores.  If you wanted to implement, say, Smalltalk or Objective-C you could use this to allow their calls to proceed in parallel.  My particular interest is based on AI.  I'm not talking about neural-nets, as that, I think, would incur too much overhead if implemented with even this kind of simplified message passing, but one where several "dumb" processes are continually running in the background and sending messages whenever they detect something interesting.  (So it would also be useful if the tasks could be assigned an adjustable priority.)

The kind of system I'm envisioning should be able to adjust to a variable number of processors...even variable while the program is running.  (Yeah, that's not what we're talking about here.  I'm talking about the higher level design.)

Erlang can probably do what I want, but it's incredibly slow.  In tests I've run it's come out even slower than Ruby, which is slower than Python.  Currently my choice is between D and Java. with a preference for D.  Java has the libraries, but D is a better design fit with what I want to do.

(I can't give detailed specifics, because I haven't yet done any programming of that part of the process.  But the rough design calls for lots of small modules with a coordinator that manages them.  I don't like having the coordinator be so central, but every threading model I've seen has a central controller.  I'd really rather do something more like Linux/Unix daemons (see the Pandemonium paper, which may have originally inspired Unix).
October 22, 2009
dsimcha wrote:

> For now, parallelFuture was designed with a single producer, multiple worker
> model.  Absolutely no attempt was made to allow for tasks running in the task pool
> to themselves submit jobs to the same task pool, because it would have made things
> more complicated and I couldn't think of any use cases.

like recursion a function is gona need to call another function, a thread needs to be able to spawn threads and task should be able to create new tasks. Making newly spawned tasks stay on the same thread is good optimization. This shouldn't need a specific use case.

> I designed the lib with
> the types of use cases I encounter in my work in mind.  (mathy, pure throughput
> oriented computing on large, embarrassingly parallel problems.)  If someone comes
> up with a compelling use case, though, I'd certainly consider adding such
> abilities provided they don't interfere with performance or API simplicity for the
> more common cases.
> 
> To make this discussion simple, let's define F1 as a future/task submitted by the
> main producer thread, and F2 as a task/future submitted by F1.  The queue is (for
> now) strictly FIFO, except that if you have a pointer to the Task/Future object
> you can steal a job.  When F1 submits F2 to the queue, F2 goes to the back of the
> queue like anything else.  This means when F1 waits on F2, it is possible to have
> a cyclical dependency (F1 waiting on F2, F2 waiting for a worker thread populated
> by F1).  This is mitigated by work stealing (F1 may just steal F2 and do it in its
> own thread).

I don't like that ^ idea of simple discussion with the many F1 and F2 all over the place. I hope this video can help visualize some ideas: http://channel9.msdn.com/pdc2008/TL26/

> 
> In parallel map and foreach, I should probably document this, but for now it's
> undefined behavior for the mapping function or parallel foreach loop body to
> submit jobs to the task pool and wait on them, and in practice will likely result
> in deadlocks.

You want to document that as undefined behavior? It can be made to work.
« First   ‹ Prev
1 2