[Comments below] Edd Dawson wrote:
Hello folks,
I'm currently trying to write tests for some threaded code. However, context switching is not predictable and scheduling is unfair on the platforms concerned, so there is no deterministic behaviour on which I can rely.
Here's a sketch of my task_queue class:
class task_queue : noncopyable { public: // use num_threads threads to churn through tasks task_queue(unsigned num_threads);
// Don't start any new tasks. // Return as soon as all active tasks are complete. ~task_queue(); // throw()
// Add a task to the back of the queue void append(const function<void ()> &task);
private: // ... };
The tasks should be picked up in FIFO order. Without some kind of deterministic threading implementation, I can't test this with the given interface. To truly test FIFO ordering, looking at when tasks start and/or finish is insufficient. Instead I need to see when a task is picked up by a thread. I can hack in some hooks, but that seems wrong(?).
So! I had this slightly wacky thought. Would it be possible to implement the boost.Thread API using Boost.Coroutine? Coroutine yield()s would be performed inside blocking calls to whatever_lock_type::lock(), thread::join(), condition::wait() and so on, perhaps; I haven't completely thought this through, yet... :)
You are entering an extremely interesting area for those of us that are truly test-infected. I've got no comments on the technical feasibility, but for certain testing scenarios the possibility of controlling "thread"-scheduling would definitely be valuable. As a pretty experienced TFD:er with threaded code, I should give you a warning though: the race conditions and deadlock sequences that you are foreseeing, and therefore writing tests for, are generally the ones that you will handle correctly. You should of course be writing tests for those as well as anything else, but there is no real replacement for running "higher-level" unit tests as well with a high amount of contention involved. Such tests are significantly slower than normal unit tests, and might become a real-world annoyance when the amount of unit tests are closing up to several hundreds or even thousands, so your testing tool should have a way of discriminating slow tests for normal compile/link/test cycles. Also, a general recommendataion is to split the code into smaller units whose logic can be separately tested without the involvement of threads before aggregating them into a larger component (as e.g. the task_queue above). What about adding a non-public task_queue_container class that contains enough observer methods to be able to verify the desired behaviour, both in single-threaded and multi-threaded contexts, and letting the public task_queue delegate most of its job to that implementation? All that being said though, I still think your idea is worthwhile pursuing. If you don't get any positive feedback from the Boost community, feel free to e-mail me directly if you ever get any further in this area. Regards, Johan Nilsson