Le 01/05/13 23:08, Vicente J. Botet Escriba a écrit :

Le 01/05/13 20:39, Niall Douglas a écrit :

...

It is just that I don't know the term. As I don't know the domain (the words) you are talking about we can stop the exchange here if you prefer, or you could continue to explaining me the things I don't know/understand.

I have just read N3388 - Using Asio with C++11 and a think I start to understand the ASIO design. Could we say the completion handlers are continuations? And that the user use to master which thread executes the continuation by calling the io_service::run() function? And that io_service can have associated also several threads (to form a thread pool). And that in order to avoid data races the user uses strands to serialize the completion handlers that have access to shared data? While continuation passing style could be well adapted to protocols (and surely it is) , there are other domains that would prefer to write his code in a more linear way. It is true that with C++11 lambdas the continuation passing style is less opaque. I hope that now I would be able to understand your concerns better. Vicente

Le 02/05/13 11:58, Dan Lincan a écrit :

Hello,

I have updated my proposal following your guide and tried to answer and give solution to most of the problems. I have tried to separate the interface with the implementation details so when choosing an interface a feature can be easily added by following the exact interface header so multiple implementations can be used without affecting the user(ex: scheduling algorithm for dynamic thereads, bounded/unbounded queues etc). I hope this approach is better.

You can find the pdf version here [1]. The previous version of the proposal can be found here [2].

Thank you very much for the feedback, suggestions, guidance and the really fast answers.

[1] http://danlincan.3owl.com/gsoc/final-proposal.pdf [2] http://danlincan.3owl.com/gsoc/Proposal.pdf

Hi Dan, *The new proposal contains:* *Waitable tasks* This feature gives the user the option to wait for a task to finish and retrieve its result. Drawback is that maybe the user is not interested in the result of the task. Interface template boost::future::type> submit(Function&&f, Args&&args); My comments: what do you think of separating the responsabilities, the thread pool schedules Callables (void()) and a submit free function manage with futures: class thread_pool { public: template<class Callable > void submit(Callable&&); ... }; template boost::future::type> submit(Executor& e, Function&& f, Args&& args); Implementing this function needs to use the same techniques than boost::async()* The new proposal contains:* *Task scheduling priority* The user would be allowed to give the tasks a priority, something similar to the operating system process priority: low, medium, high etc. Performance is lost when the algorithm to schedule tasks is running. Based on the implementation, this can be more that O(1), thus big overhead. Another problem is fair scheduling. Low priority tasks have to run at some point even if only high priority tasks are being submitted to the threadpool. Interface template boost::future::type> submit(Function&&f, Args&&args, Priorityp = default_priority()); My comments: I don't think adding priorities in this way would improve the performances. *The new proposal contains:* *Task execution order* For example the user might want x tasks to be executed, then only after all x have been executed another y can be executed and so on. So the user could assign the priority 1 to the x tasks and 2 to y tasks. The threadpool will know that level 2 tasks can only be executed after level 1 tasks are done. Performance is lost when a task is chosen to be executed. For example, if using priority queues, the complexity would be O(logn) plus the locks which can slow down a lot the scheduling. This feature can be used as a kind of serial executor by giving higher and higher numbers to the order argument. Interface template boost::future::type> *submit**(**Function**&&***f*, **Args&&*args*, **Order*o= default_order()); *Callback functions / Continuations* The user can also set a callback function when submitting a task to the threadpool. This function will run as soon as the task is finished by the threadpool. Drawback: Since the callback function is runned by the thread which executs the task, this thread can be block if the callback function is blocking. Also the callback has no parameters. Interface template boost::future::type> *submit**(**Function**&&***f*, **Args&&*args*, **Callback*c= default_callback()); My comments: I prefer continuations than callbacks and prefer an interface that explicitly states the dependencies. when_all(submit(f), submit(g)).then(h); *The new proposal contains:* *Task cancellation* With this feature the use can cancel a submitted task. If the task is currently being executed it will be interrupted. For implementation boost::thread::intterupts will be activated. template boost::future::type> *submit**(**Function**&&***f*, **Args&&*args*, ****Task&*task); Task has the function cancel which cancels/interrupts when called. My comments: This is a little bit complex from the users point of view. I guess the best is to either: * adding to boost::future the possibility to interrupt the task, or * returning a interrupted_future<T> or task<T> that can be interrupted. There is an additional requirement. The user must ensure that all the tasks related to some object are executed sequentially but the task submission don't knows about the pending tasks. Continuations don't solve the problem. What would you add to make the users life easier? (please read the discussion about threadpool and ASIO, the answer is someway there). Best, Vicente

*The new proposal contains:*

*Waitable tasks*

This feature gives the user the option to wait for a task to finish and retrieve its result.

Drawback is that maybe the user is not interested in the result of the task.

Interface

template

boost::future::type>

submit(Function&&f, Args&&args);

My comments:

what do you think of separating the responsabilities, the thread pool schedules Callables (void()) and a submit free function manage with futures:

class thread_pool { public:

template<class Callable > void submit(Callable&&); ... };

template boost::future::type> submit(Executor& e, Function&& f, Args&& args);

Implementing this function needs to use the same techniques than boost::async()*

What techniques from boost::async() are you referring to?

The new proposal contains:*

*Task scheduling priority*

The user would be allowed to give the tasks a priority, something similar to the operating system process priority: low, medium, high etc.

Performance is lost when the algorithm to schedule tasks is running. Based on the implementation, this can be more that O(1), thus big overhead. Another problem is fair scheduling. Low priority tasks have to run at some point even if only high priority tasks are being submitted to the threadpool.

Interface

template

boost::future::type>

submit(Function&&f, Args&&args, Priorityp = default_priority());

My comments:

I don't think adding priorities in this way would improve the performances.

Well I thought about adding priorities this way to achieve close to O(1) performance. I had in mind something similar to counting sort. As long as we know that we have at most x priorities level ( in this case x=3: low, medium, high ), we can store the tasks in an array of containers(queue/lock-free queue/others) so submitting a task to the threadpool can be done in O(1). The only problem after this is having to deal with fair scheduling. This can be done in O(1) too so O(1) performance overall.

*The new proposal contains:*

*Task execution order*

For example the user might want x tasks to be executed, then only after all x have been executed another y can be executed and so on. So the user could assign the priority 1 to the x tasks and 2 to y tasks. The threadpool will know that level 2 tasks can only be executed after level 1 tasks are done.

Performance is lost when a task is chosen to be executed. For example, if using priority queues, the complexity would be O(logn) plus the locks which can slow down a lot the scheduling.

This feature can be used as a kind of serial executor by giving higher and higher numbers to the order argument.

Interface

template

boost::future::type>

*submit**(**Function**&&***f*, **Args&&*args*, **Order*o= default_order());

*Callback functions / Continuations*

The user can also set a callback function when submitting a task to the threadpool. This function will run as soon as the task is finished by the threadpool.

Drawback: Since the callback function is runned by the thread which executs the task, this thread can be block if the callback function is blocking. Also the callback has no parameters.

Interface

template

boost::future::type>

*submit**(**Function**&&***f*, **Args&&*args*, **Callback*c= default_callback());

My comments:

I prefer continuations than callbacks and prefer an interface that explicitly states the dependencies.

when_all(submit(f), submit(g)).then(h);

This is what I had in mind with this feature. I should have been more clear as this is the reason I have added [1] as reference.

*The new proposal contains:*

*Task cancellation*

With this feature the use can cancel a submitted task. If the task is currently being executed it will be interrupted. For implementation boost::thread::intterupts will be activated.

template

boost::future::type>

*submit**(**Function**&&***f*, **Args&&*args*, ****Task&*task);

Task has the function cancel which cancels/interrupts when called.

My comments:

This is a little bit complex from the users point of view. I guess the best is to either: * adding to boost::future the possibility to interrupt the task, or * returning a interrupted_future<T> or task<T> that can be interrupted.

Then I would prefer the first alternative, extending boost::future. For the user it would be really nice to use just result.interrupt() where result is a future.

There is an additional requirement. The user must ensure that all the tasks related to some object are executed sequentially but the task submission don't knows about the pending tasks. Continuations don't solve the problem. What would you add to make the users life easier? (please read the discussion about threadpool and ASIO, the answer is someway there).

I will search it and get back with an answer. [1] http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2013/n3558.pdf Regards, Dan

Dan wrote on Thursday, May 2, 2013 at 13:58:35:

Hello,

I have updated my proposal following your guide and tried to answer and give solution to most of the problems. I have tried to separate the interface with the implementation details so when choosing an interface a feature can be easily added by following the exact interface header so multiple implementations can be used without affecting the user(ex: scheduling algorithm for dynamic thereads, bounded/unbounded queues etc). I hope this approach is better.

You can find the pdf version here [1]. The previous version of the proposal can be found here [2].

Thank you very much for the feedback, suggestions, guidance and the really fast answers.

[1] http://danlincan.3owl.com/gsoc/final-proposal.pdf [2] http://danlincan.3owl.com/gsoc/Proposal.pdf

perhaps not _this_ time, that is not this 2013 GSOC, but eventually you may want to consider a lock-free priority queue algorithm for storing tasks and not use locks although lock-free algorithms are somewhat hard to comprehend, all the more so about _priority_ queues, but in the end that's, beside other things, what one would expect from modern implementation of thread pool moreover, you, as a student loving algorithms and analysis, may like to dig into the lock-free algorithms in general good luck -- Pavel P.S. if you notice a grammar mistake or weird phrasing in my message please point it out

Vicente wrote on Thursday, May 2, 2013 at 22:40:06:

...

...

Thank you very much for the feedback, suggestions, guidance and the really fast answers. [1] http://danlincan.3owl.com/gsoc/final-proposal.pdf [2] http://danlincan.3owl.com/gsoc/Proposal.pdf

Dan wrote on Thursday, May 2, 2013 at 13:58:35: perhaps not _this_ time, that is not this 2013 GSOC, but eventually you may want to consider a lock-free priority queue algorithm for storing tasks and not use locks It was in my mind to reuse Boos.LookFree queue. Why do you need a

Le 02/05/13 17:24, pavel a écrit : priority queue here?

dan write in his proposal (the first link above) about task scheduling priority and that the implementation may require a priority queue with locks i just pointed out that it will be interesting to use particularly lock-free priority queue for that purpose instead of locking algorithm certainly, if boost.lockfree already had priority queue, one must have been (re)used that implementation, and in fact it's a pity that boost.lockfree hasn't one (as well as some kind of map/set) as for priority task scheduling -- i consider it an essential feature of a threadpool facility that certainly should be there and yes, boost absolutely needs a threadpool implementation i hope this project will be a good start -- Pavel P.S. if you notice a grammar mistake or weird phrasing in my message please point it out

Le 02/05/13 11:58, Dan Lincan a écrit :

Hello,

I have updated my proposal following your guide and tried to answer and give solution to most of the problems. I have tried to separate the interface with the implementation details so when choosing an interface a feature can be easily added by following the exact interface header so multiple implementations can be used without affecting the user(ex: scheduling algorithm for dynamic thereads, bounded/unbounded queues etc). I hope this approach is better.

You can find the pdf version here [1]. The previous version of the proposal can be found here [2].

Thank you very much for the feedback, suggestions, guidance and the really fast answers.

[1] http://danlincan.3owl.com/gsoc/final-proposal.pdf [2] http://danlincan.3owl.com/gsoc/Proposal.pdf

Dan, are you there? Vicente