Return value or Exception or something else ?

Mr Brainley

This is my function :

CThreadPool::t_Job& CThreadPool::CJobQueue::getJob()
{
WaitForSingleObject(m_JobListMutex, INFINITE);

for ( std::deque::iterator it = m\_JobList.begin(); it != m\_JobList.end(); it++ )
	if ( it->m\_JobState != JOB\_WORKING)
	{
		it->m\_JobState	= JOB\_WORKING;
		ReleaseMutex(m\_JobListMutex);
		return (\*it);
	}

ReleaseMutex(m\_JobListMutex);
//	Need to do something here to indicate no item has been found

}

I don't want to use a pointer for return. And i don't think it's apropriate to throw an exception, since that scenario is unavoidable, in fact quite regular. My only Idea left would be to return some special reference to a static element that functions as sort of a NULL. Any other ideas/comments ?

Waldermort

why not use an enum type of predefined statuses?

led mike

Mr.Brainley wrote:

I don't want to use a pointer for return.

Why not? I also question why INFINITE is hard coded? Does not seem appropriate for thread pooling. Also does your pool provide a timeout mechanism to account for user code hogging and hanging threads?

Mr.Brainley wrote:

Any other ideas

A smart pointer implementation that would encapsulate returning the thread to the pool etc.

led mike

Jorgen Sigvardsson

Return pointer instead of reference? That way you can safely use NULL to indicate "no job". Using exceptions for that is in my opinion a complete waste of resources... May I ask why you are using mutexes? You could speed things up by using spinlocks (CRITICAL_SECTION). See docs for EnterCriticalSection(), LeaveCriticalSection(), etc, in the MSDN library. I'm assuming that the mutexes are never, or very seldom, held for a longer period of time. If they are, then mutexes is preferable, because they will put waiting thread to sleep rather than leaving it to "spin".

-- Now with chucklelin

Mr Brainley

I think my thread pool works different from what you think. Basically, i have a central semaphore that is increased by one each time a Job is added. On that, a thread wakes up and asks the queue for a job. The t_Job type of my example is a simple class containing a pointer to an interface with a function to execute, and state-information. The INFINITE wait here is just for the mutex on the jobqueue, wich of course needs to be synchronized. The pool does not have a timeout mechanism. I don't think it is neccessary, since i write it for a very special purpose, but anyhow i would'nt know how to implement that. Can you give me a hint there ? The main idea was, that whenever is thread is wakened by the semaphore, that means that there actually IS a job. But in Win32 you have to give a maximum count for a semaphore. Since i don't want to limit the number of jobs in my queue, i tried to find another solution, wich brought me to the problem at hand. I think i will just live with the limit (can set it high). The maximum workloud can be estimated and that will have to suffice.

Jorgen Sigvardsson

I see now that you are opposing the use of pointers.. why? Pointers are an intrinsic concept in C++ - why work against it? :~

-- Presented in doublevision (where drunk)

Mr Brainley

Because pointers are evil. Check this FAQ for more on that topic. Also, i couldn't get it to work. Can't cast the iterator to a pointer. No idea how it can be done.

Mark Salsbery

Have you ever used an IO completion port? It can be used for what you are doing without the semaphore release count limit. Instead of IO you would use PostQueuedCompletionStatus() to post jobs and your pool of threads wait for jobs using GetQueuedCompletionStatus(). When a waiting thread gets a completion packet it already has the job (pointer) so it doen't need to retrieve it from a list. It works pretty efficiently. Just a thought. Mark

Mr Brainley

Joergen Sigvardsson wrote:

Pointers are an intrinsic concept in C++

Actually, pointers are an intrisic concept in C. C++ has references, and that has a reason. I try to follow that guideline.

Mr Brainley

I was not aware of such a thing. Thanks alot !

Mr Brainley

Thanks for the advice on the Mutex. You are right, they are never held for long. I will change that.

Mark Salsbery

Mr.Brainley wrote:

I was not aware of such a thing. Thanks alot !

No problem. I just happen to be working on some code using one right now :) There's only 3 APIs and 1 structure involved so it's pretty simple and powerful. You'll see most articles describe using them for scalable multithreaded IO but they also make a handy queue for thread pools. Cheers, Mark

Jorgen Sigvardsson

Uhm.. that is a rather dull guideline. Just because some academic has decided that references are better than pointers (which the majority of the C++ community uses), it's counter productive NOT to use them in cases like this, as it would cut your development by at least an hour (seeing that you posted this approximately an hour ago). But hey, it's your time. :)

-- Based on a True Story

Mr Brainley

Ok, you're right. And i don't do that just because some academic decided it. I use pointers often enough. But i also take my time to explore other solutions, since i think i still have much to learn. That is why i don't just use pointers because everyone else uses them. There may be a better solution. ( and remember : The fast path leads to the dark side ;)

Mr Brainley

actually, I'm implementing this thread pool for a scalable multithreaded IO application. Well, seems i don't need to do that anymore.

Mark Salsbery

Then you'll really like IO completion ports then. You can use the same pool of threads for handling queued jobs and queued IO requests if you choose to. Works slick!

Zac Howland

You should wrap your mutex calls in a class (open it in the constructor and release it in the destructor). This actually allows for the exception option (otherwise, throwing an exception would be bad since the code would not be exception safe). Given what you are trying to do, I would probably pass in a reference to be returned to as an argument and return a bool or some other status indicator as a return value. It would make your code look something like this:

class MutexHolder
{
public:
	MutexHolder(HANDLE mutex)
	{
		_Mutex = mutex;
		WaitForSingleObject(_Mutex, INFINITE);
		// I believe this should actually be OpenMutex
		// In which case, you would change the constructor arguement to a string
	}

	~MutexHolder()
	{
		ReleaseMutex(_Mutex);
	}

private:
	HANDLE _Mutex;
};

bool job_not_working(const your_deque_type& t)
{
	if (t.m_JobState != JOB_WORKING)
	{
		return true;
	}
}

bool CThreadPool::CJobQueue::getJob(CThreadPool::t_Job& job)
{
	MutexHolder(m_JobListMutex);

	deque<your_deque_type>::iterator it = find_if(m_JobList.begin(), m_JobList.end(), job_not_working);
	if (it != m_JobList.end())
	{
		job = *it;
		job.m_JobState = JOB_WORKING;
		return true;
	}
	return false;
}

If you decide to become a software engineer, you are signing up to have a 1/2" piece of silicon tell you exactly how stupid you really are for 8 hours a day, 5 days a week Zac

Jorgen Sigvardsson

The problem with that approach, is that you'd be copying the job object. That may not always be possible. If you're going to return a status value, I'd suggest the COM way of returning the job: pointer to pointer, or reference to pointer (if you're hell bent on using references)

-- From the Makers of Futurama

Zac Howland

The object is stored in a deque, therefore it is already being copied. The only way to avoid that is to have the deque store pointers, in which case the approach still works, you just need to pass in a reference to a pointer instead.

If you decide to become a software engineer, you are signing up to have a 1/2" piece of silicon tell you exactly how stupid you really are for 8 hours a day, 5 days a week Zac

Zac Howland

Joergen Sigvardsson wrote:

Uhm.. that is a rather dull guideline. Just because some academic has decided that references are better than pointers (which the majority of the C++ community uses), it's counter productive NOT to use them in cases like this, as it would cut your development by at least an hour (seeing that you posted this approximately an hour ago).

Knowing how to use references increases code readability, maintainability, and decreases the liklihood of bugs being introduced via improper usage of pointers. In general, if you can get away with not using pointers, you should do so.

If you decide to become a software engineer, you are signing up to have a 1/2" piece of silicon tell you exactly how stupid you really are for 8 hours a day, 5 days a week Zac