Jump to content

  • Log In with Google      Sign In   
  • Create Account


Tools for proving things for multithreaded systems


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
6 replies to this topic

#1 PAndersson   Members   -  Reputation: 394

Like
0Likes
Like

Posted 09 February 2013 - 06:50 PM

I have finalized the intiial threading model for a program I'm working on, and due to the nature of its enviroment I have to allow deadlocks to occur or loose much of the benift from multithreading. Fortunatly, it also enables easy detection and resolution of deadlocks without rolling anything back; except for one special case though the programmer has to go out of his way to make that one possible and those can still easily be detected though not resolved.

 

Anyway, I do not think there are any more cases where deadlocking can occur but I'm not 100% certain either. So I'm wondering if there are any good tools or methods for proving that deadlocking is impossible (other than that onbe case I mentioned) in a system of unknown size? The only method I'm aware of is petri nets, though I do not know if that method can be adapted to work with a model of unknown size or not.



Sponsor:

#2 KulSeran   Members   -  Reputation: 2163

Like
0Likes
Like

Posted 10 February 2013 - 12:44 AM

You might be looking for tools like: http://research.microsoft.com/en-us/projects/CHESS/
But I think the better question is why does your threading model require this?

#3 Hodgman   Moderators   -  Reputation: 27523

Like
1Likes
Like

Posted 10 February 2013 - 02:16 AM

Intel Thread checker can watch you application and report any code that represents a potential danger, but it has to do this at runtime instead of as a static-analysis. Also, it costs money.

 

I've the same feeling as KulSeran though -- the fact that your strategy is at risk of deadlocks and has to implement detection/resolution sounds very... interesting.



#4 PAndersson   Members   -  Reputation: 394

Like
0Likes
Like

Posted 10 February 2013 - 03:30 AM

I went with this solution as it minimizes locking and ironically, the risk of deadlocks. The source of my issue is that openGL requires every task for a given openGL context to be performed in the same thread, this leads to tasks depending on other tasks for completion. Sometimes these tasks may end up in the same task queue and in the wrong order. The solution I have now simply finds out if a task waits for another in the same queue, and lets the one that is being waited for skip ahead.

 

Had the requirement to do certain tasks in a given thread not been there, task inter-dependencies would have disappeared and this problem would not be there.



#5 Shannon Barber   Moderators   -  Reputation: 1355

Like
0Likes
Like

Posted 11 February 2013 - 11:02 PM

I really wish they would do something about this issue in OGL.

 

Why do your tasks end-up out-of-order?

(I used what I think is called a 'monitor' to arbitrate access to the context.)


- The trade-off between price and quality does not exist in Japan. Rather, the idea that high quality brings on cost reduction is widely accepted.-- Tajima & Matsubara

#6 phantom   Moderators   -  Reputation: 6785

Like
0Likes
Like

Posted 12 February 2013 - 03:38 AM

I really wish they would do something about this issue in OGL.

To be honest, this is the least of OGL's problems ;)

Personally I would generate a series of command lists on the different threads and then submit them via one single submission thread rather than mess about with trying to migrate the GL context between threads in a reliable manner. It'll probably work out faster as constant calls into GL from a single thread with nicely arranged data is more likely to hit instruction and data caches than jumping between threads, and thus potentially cores, and the associcated release/obtain code required to deal with it.

#7 Kylotan   Moderators   -  Reputation: 3329

Like
2Likes
Like

Posted 12 February 2013 - 10:23 AM

Yeah, this arguably isn't a 'problem' with OpenGL - lots of systems are essentially going to require serialised access to a resource and it's your job to ensure that the series of actions is in the correct order. I would humbly suggest that splitting serialised tasks into multiple queues only to make them block each other based on task ordering is probably the wrong way to go about using task queues.

 

Had the requirement to do certain tasks in a given thread not been there, task inter-dependencies would have disappeared and this problem would not be there.

 

That doesn't make sense to me - if the only dependency between tasks is that only one can execute at a time then the ordering issue you mention wouldn't exist. And if there is an ordering issue, then you need to resolve that before attempting to parallelise the tasks because once they're in separate queues, ordering between the queues no longer applies. (Without adding in some sort of global lock which mostly destroys the purpose of having multiple queues.)






Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS