I have a queue with elements which needs to be processed. I want to process these elements in parallel. The will be some sections on each element which need to be synchronized. At any point in time there can be max num_threads running threads.
I'll provide a template to give you an idea of what I want to achieve.
queue q
process_element(e)
{
lock()
some synchronized area
// a matrix access performed here so a spin lock would do
unlock()
...
unsynchronized area
...
if( condition )
{
new_element = generate_new_element()
q.push(new_element) // synchonized access to queue
}
}
process_queue()
{
while( elements in q ) // algorithm is finished condition
{
e = get_elem_from_queue(q) // synchronized access to queue
process_element(e)
}
}
I can use
pthreads
openmp
intel thread building blocks
Top problems I have
Make sure that at any point in time I have max num_threads running threads
Lightweight synchronization methods to use on queue
My plan is to the intel tbb concurrent_queue for the queue container. But then, will I be able to use pthreads functions ( mutexes, conditions )? Let's assume this works ( it should ). Then, how can I use pthreads to have max num_threads at one point in time? I was thinking to create the threads once, and then, after one element is processes, to access the queue and get the next element. However it if more complicated because I have no guarantee that if there is not element in queue the algorithm is finished.
My question
Before I start implementing I'd like to know if there is an easy way to use intel tbb or pthreads to obtain the behaviour I want? More precisely processing elements from a queue in parallel
Note: I have tried to use tasks but with no success.
First off, pthreads gives you portability which is hard to walk away from. The following appear to be true from your question - let us know if these aren't true because the answer will then change:
1) You have a multi-core processor(s) on which you're running the code
2) You want to have no more than num_threads threads because of (1)
Assuming the above to be true, the following approach might work well for you:
Create num_threads pthreads using pthread_create
Optionally, bind each thread to a different core
q.push(new_element) atomically adds new_element to a queue. pthreads_mutex_lock and pthreads_mutex_unlock can help you here. Examples here: http://pages.cs.wisc.edu/~travitch/pthreads_primer.html
Use pthreads_mutexes for dequeueing elements
Termination is tricky - one way to do this is to add a TERMINATE element to the queue, which upon dequeueing, causes the dequeuer to queue up another TERMINATE element (for the next dequeuer) and then terminate. You will end up with one extra TERMINATE element in the queue, which you can remove by having a named thread dequeue it after all the threads are done.
Depending on how often you add/remove elements from the queue, you may want to use something lighter weight than pthread_mutex_... to enqueue/dequeue elements. This is where you might want to use a more machine-specific construct.
TBB is compatible with other threading packages.
TBB also emphasizes scalability. So when you port over your program to from a dual core to a quad core you do not have to adjust your program. With data parallel programming, program performance increases (scales) as you add processors.
Cilk Plus is also another runtime that provides good results.
www.cilkplus.org
Since pThreads is a low level theading library you have to decide how much control you need in your application because it does offer flexibility, but at a high cost in terms of programmer effort, debugging time, and maintenance costs.
My recommendation is to look at tbb::parallel_do. It was designed to process elements from a container in parallel, even if the container itself is not concurrent; i.e. parallel_do works with an std::queue correctly without any user synchronization (of course you would still need to protect your matrix access inside process_element(). Moreover, with parallel_do you can add more work on the fly, which looks like what you need, as process_element() creates and adds new elements to the work queue (the only caution is that the newly added work will be processed immediately, unlike putting in a queue which would postpone processing till after all "older" items). Also, you don't have to worry about termination: parallel_do will complete automatically as soon as all initial queue items and new items created on the fly are processed.
However, if, besides the computation itself, the work queue can be concurrently fed from another source (e.g. from an I/O processing thread), then parallel_do is not suitable. In this case, it might make sense to look at parallel_pipeline or, better, the TBB flow graph.
Lastly, an application can control the number of active threads with TBB, though it's not a recommended approach.
Related
Anecdotally, I've found that a lot of programmers mistakenly believe that "lock-free" simply means "concurrent programming without mutexes". Usually, there's also a correlated misunderstanding that the purpose of writing lock-free code is for better concurrent performance. Of course, the correct definition of lock-free is actually about progress guarantees. A lock-free algorithm guarantees that at least one thread is able to make forward progress regardless of what any other threads are doing.
This means a lock-free algorithm can never have code where one thread is depending on another thread in order to proceed. E.g., lock-free code can not have a situation where Thread A sets a flag, and then Thread B keeps looping while waiting for Thread A to unset the flag. Code like that is basically implementing a lock (or what I would call a mutex in disguise).
However, other cases are more subtle and there are some cases where I honestly can't really tell if an algorithm qualifies as lock-free or not, because the notion of "making progress" sometimes appears subjective to me.
One such case is in the (well-regarded, afaik) concurrency library, liblfds. I was studying the implementation of a multi-producer/multi-consumer bounded queue in liblfds - the implementation is very straightforward, but I cannot really tell if it should qualify as lock-free.
The relevant algorithm is in lfds711_queue_bmm_enqueue.c. Liblfds uses custom atomics and memory barriers, but the algorithm is simple enough for me to describe in a paragraph or so.
The queue itself is a bounded contiguous array (ringbuffer). There is a shared read_index and write_index. Each slot in the queue contains a field for user-data, and a sequence_number value, which is basically like an epoch counter. (This avoids ABA issues).
The PUSH algorithm is as follows:
Atomically LOAD the write_index
Attempt to reserve a slot in the queue at write_index % queue_size using a CompareAndSwap loop that attempts to set write_index to write_index + 1.
If the CompareAndSwap is successful, copy the user data into the
reserved slot.
Finally, update the sequence_index on the
slot by making it equal to write_index + 1.
The actual source code uses custom atomics and memory barriers, so for further clarity about this algorithm I've briefly translated it into (untested) standard C++ atomics for better readability, as follows:
bool mcmp_queue::enqueue(void* data)
{
int write_index = m_write_index.load(std::memory_order_relaxed);
for (;;)
{
slot& s = m_slots[write_index % m_num_slots];
int sequence_number = s.sequence_number.load(std::memory_order_acquire);
int difference = sequence_number - write_index;
if (difference == 0)
{
if (m_write_index.compare_exchange_weak(
write_index,
write_index + 1,
std::memory_order_acq_rel
))
{
break;
}
}
if (difference < 0) return false; // queue is full
}
// Copy user-data and update sequence number
//
s.user_data = data;
s.sequence_number.store(write_index + 1, std::memory_order_release);
return true;
}
Now, a thread that wants to POP an element from the slot at read_index will not be able to do so until it observes that the slot's sequence_number is equal to read_index + 1.
Okay, so there are no mutexes here, and the algorithm likely performs well (it's only a single CAS for PUSH and POP), but is this lock-free? The reason it's unclear to me is because the definition of "making progress" seems murky when there is the possibility that a PUSH or POP can always just fail if the queue is observed to be full or empty.
But what's questionable to me is that the PUSH algorithm essentially reserves a slot, meaning that the slot can never be POP'd until the push thread gets around to updating the sequence number. This means that a POP thread that wants to pop a value depends on the PUSH thread having completed the operation. Otherwise, the POP thread will always return false because it thinks the queue is EMPTY. It seems debatable to me whether this actually falls within the definition of "making progress".
Generally, truly lock-free algorithms involve a phase where a pre-empted thread actually tries to ASSIST the other thread in completing an operation. So, in order to be truly lock-free, I would think that a POP thread that observes an in-progress PUSH would actually need to try and complete the PUSH, and then only after that, perform the original POP operation. If the POP thread simply returns that the queue is EMPTY when a PUSH is in progress, the POP thread is basically blocked until the PUSH thread completes the operation. If the PUSH thread dies, or goes to sleep for 1,000 years, or otherwise gets scheduled into oblivion, the POP thread can do nothing except continuously report that the queue is EMPTY.
So does this fit the defintion of lock-free? From one perspective, you can argue that the POP thread can always make progress, because it can always report that the queue is EMPTY (which is at least some form of progress I guess.) But to me, this isn't really making progress, since the only reason the queue is observed as empty is because we are blocked by a concurrent PUSH operation.
So, my question is: is this algorithm truly lock-free? Or is the index reservation system basically a mutex in disguise?
This queue data structure is not strictly lock-free by what I consider the most reasonable definition. That definition is something like:
A structure is lock-free if only if any thread can be indefinitely
suspended at any point while still leaving the structure usable by the
remaining threads.
Of course this implies a suitable definition of usable, but for most structures this is fairly simple: the structure should continue to obey its contracts and allow elements to be inserted and removed as expected.
In this case a thread that has succeeded in incrementing m_write_increment, but hasn't yet written s.sequence_number leaves the container in what will soon be an unusable state. If such a thread is killed, the container will eventually report both "full" and "empty" to push and pop respectively, violating the contract of a fixed size queue.
There is a hidden mutex here (the combination of m_write_index and the associated s.sequence_number) - but it basically works like a per-element mutex. So the failure only becomes apparent to writers once you've looped around and a new writer tries to get the mutex, but in fact all subsequent writers have effectively failed to insert their element into the queue since no reader will ever see it.
Now this doesn't mean this is a bad implementation of a concurrent queue. For some uses it may behave mostly as if it was lock free. For example, this structure may have most of the useful performance properties of a truly lock-free structure, but at the same time it lacks some of the useful correctness properties. Basically the term lock-free usually implies a whole bunch of properties, only a subset of which will usually be important for any particular use. Let's look at them one by one and see how this structure does. We'll broadly categorize them into performance and functional categories.
Performance
Uncontended Performance
The uncontended or "best case" performance is important for many structures. While you need a concurrent structure for correctness, you'll usually still try to design your application so that contention is kept to a minimum, so the uncontended cost is often important. Some lock-free structures help here, by reducing the number of expensive atomic operations in the uncontended fast-path, or avoiding a syscall.
This queue implementation does a reasonable job here: there is only a single "definitely expensive" operation: the compare_exchange_weak, and a couple of possibly expensive operations (the memory_order_acquire load and memory_order_release store)1, and little other overhead.
This compares to something like std::mutex which would imply something like one atomic operation for lock and another for unlock, and in practice on Linux the pthread calls have non-negligible overhead as well.
So I expect this queue to perform reasonably well in the uncontended fast-path.
Contended Performance
One advantage of lock-free structures is that they often allow better scaling when a structure is heavily contended. This isn't necessarily an inherent advantage: some lock-based structures with multiple locks or read-write locks may exhibit scaling that matches or exceeds some lock-free approaches, but it is usually that case that lock-free structures exhibit better scaling that a simple one-lock-to-rule-them-all alternative.
This queue performs reasonably in this respect. The m_write_index variable is atomically updated by all readers and will be a point of contention, but the behavior should be reasonable as long as the underlying hardware CAS implementation is reasonable.
Note that a queue is generally a fairly poor concurrent structure since inserts and removals all happen at the same places (the head and the tail), so contention is inherent in the definition of the structure. Compare this to a concurrent map, where different elements have no particular ordered relationship: such a structure can offer efficient contention-free simultaneous mutation if different elements are being accessed.
Context-switch Immunity
One performance advantage of lock-free structures that is related to the core definition above (and also to the functional guarantees) is that a context switch of a thread which is mutating the structure doesn't delay all the other mutators. In a heavily loaded system (especially when runnable threads >> available cores), a thread may be switched out for hundreds of milliseconds or seconds. During this time, any concurrent mutators will block and incur additional scheduling costs (or they will spin which may also produce poor behavior). Even though such "unluckly scheduling" may be rare, when it does occur the entire system may incur a serious latency spike.
Lock-free structures avoid this since there is no "critical region" where a thread can be context switched out and subsequently block forward progress by other threads.
This structure offers partial protection in this area — the specifics of which depend on the queue size and application behavior. Even if a thread is switched out in the critical region between the m_write_index update and the sequence number write, other threads can continue to push elements to the queue as long as they don't wrap all the way around to the in-progress element from the stalled thread. Threads can also pop elements, but only up to the in-progress element.
While the push behavior may not be a problem for high-capacity queues, the pop behavior can be a problem: if the queue has a high throughput compared to the average time a thread is context switched out, and the average fullness, the queue will quickly appear empty to all consumer threads, even if there are many elements added beyond the in-progress element. This isn't affected by the queue capacity, but simply the application behavior. It means that the consumer side may completely stall when this occurs. In this respect, the queue doesn't look very lock-free at all!
Functional Aspects
Async Thread Termination
On advantage of lock-free structures it they are safe for use by threads that may be asynchronously canceled or may otherwise terminate exceptionally in the critical region. Cancelling a thread at any point leaves the structure is a consistent state.
This is not the case for this queue, as described above.
Queue Access from Interrupt or Signal
A related advantage is that lock-free structures can usually be examined or mutated from an interrupt or signal. This is useful in many cases where an interrupt or signal shares a structure with regular process threads.
This queue mostly supports this use case. Even if the signal or interrupt occurs when another thread is in the critical region, the asynchronous code can still push an element onto the queue (which will only be seen later by consuming threads) and can still pop an element off of the queue.
The behavior isn't as complete as a true lock-free structure: imagine a signal handler with a way to tell the remaining application threads (other than the interrupted one) to quiesce and which then drains all the remaining elements of the queue. With a true lock-free structure, this would allow the signal handler to full drain all the elements, but this queue might fail to do that in the case a thread was interrupted or switched out in the critical region.
1 In particular, on x86, this will only use an atomic operation for the CAS as the memory model is strong enough to avoid the need for atomics or fencing for the other operations. Recent ARM can do acquire and release fairly efficiently as well.
I am the author of liblfds.
The OP is correct in his description of this queue.
It is the single data structure in the library which is not lock-free.
This is described in the documentation for the queue;
http://www.liblfds.org/mediawiki/index.php?title=r7.1.1:Queue_%28bounded,_many_producer,_many_consumer%29#Lock-free_Specific_Behaviour
"It must be understood though that this is not actually a lock-free data structure."
This queue is an implementation of an idea from Dmitry Vyukov (1024cores.net) and I only realised it was not lock-free while I was making the test code work.
By then it was working, so I included it.
I do have some thought to remove it, since it is not lock-free.
Most of the time people use lock-free when they really mean lockless. lockless means a data-structure or algorithm that does not use locks, but there is no guarantee for forward progress. Also check this question. So the queue in liblfds is lockless, but as BeeOnRope mentioned is not lock-free.
A thread that calls POP before the next update in sequence is complete is NOT "effectively blocked" if the POP call returns FALSE immediately. The thread can go off and do something else. I'd say that this queue qualifies as lock-free.
However, I wouldn't say that it qualifies as a "queue" -- at least not the kind of queue that you could publish as a queue in a library or something -- because it doesn't guarantee a lot of the behaviors that you can normally expect from a queue. In particular, you can PUSH and element and then try and FAIL to POP it, because some other thread is busy pushing an earlier item.
Even so, this queue could still be useful in some lock-free solutions for various problems.
For many applications, however, I would worry about the possibility for consumer threads to be starved while a producer thread is pre-empted. Maybe liblfds does something about that?
"Lock-free" is a property of the algorithm, which implements some functionality. The property doesn't correlate with a way, how given functionality is used by a program.
When talk about mcmp_queue::enqueue function, which returns FALSE if underlying queue is full, its implementation (given in the question post) is lock-free.
However, implementing mcmp_queue::dequeue in lock-free manner would be difficult. E.g., this pattern is obviously not-lock free, as it spins on the variable changed by other thread:
while(s.sequence_number.load(std::memory_order_acquire) == read_index);
data = s.user_data;
...
return data;
I did formal verification on this same code using Spin a couple years ago for a course in concurrency testing and it is definitely not lock-free.
Just because there is no explicit "locking", doesn't mean it's lock-free. When it comes to reasoning about progress conditions, think of it from an individual thread's perspective:
Blocking/locking: if another thread gets descheduled and this can block my progress, then it is blocking.
Lock-free/non-blocking: if I am able to eventually make progress in the absence of contention from other threads, then it is at most lock-free.
If no other thread can block my progress indefinitely, then it is wait-free.
I recently heard new c++ standard features which are:
std::latch
std::barrier
I cannot figure it out ,in which situations that they are applicable and useful over one-another.
If someone can raise an example for how to use each one of them wisely it would be really helpful.
Very short answer
They're really aimed at quite different goals:
Barriers are useful when you have a bunch of threads and you want to synchronise across of them at once, for example to do something that operates on all of their data at once.
Latches are useful if you have a bunch of work items and you want to know when they've all been handled, and aren't necessarily interested in which thread(s) handled them.
Much longer answer
Barriers and latches are often used when you have a pool of worker threads that do some processing and a queue of work items that is shared between. It's not the only situation where they're used, but it is a very common one and does help illustrate the differences. Here's some example code that would set up some threads like this:
const size_t worker_count = 7; // or whatever
std::vector<std::thread> workers;
std::vector<Proc> procs(worker_count);
Queue<std::function<void(Proc&)>> queue;
for (size_t i = 0; i < worker_count; ++i) {
workers.push_back(std::thread(
[p = &procs[i], &queue]() {
while (auto fn = queue.pop_back()) {
fn(*p);
}
}
));
}
There are two types that I have assumed exist in that example:
Proc: a type specific to your application that contains data and logic necessary to process work items. A reference to one is passed to each callback function that's run in the thread pool.
Queue: a thread-safe blocking queue. There is nothing like this in the C++ standard library (somewhat surprisingly) but there are a lot of open-source libraries containing them e.g. Folly MPMCQueue or moodycamel::ConcurrentQueue, or you can build a less fancy one yourself with std::mutex, std::condition_variable and std::deque (there are many examples of how to do this if you Google for them).
Latch
A latch is often used to wait until some work items you push onto the queue have all finished, typically so you can inspect the result.
std::vector<WorkItem> work = get_work();
std::latch latch(work.size());
for (WorkItem& work_item : work) {
queue.push_back([&work_item, &latch](Proc& proc) {
proc.do_work(work_item);
latch.count_down();
});
}
latch.wait();
// Inspect the completed work
How this works:
The threads will - eventually - pop the work items off of the queue, possibly with multiple threads in the pool handling different work items at the same time.
As each work item is finished, latch.count_down() is called, effectively decrementing an internal counter that started at work.size().
When all work items have finished, that counter reaches zero, at which point latch.wait() returns and the producer thread knows that the work items have all been processed.
Notes:
The latch count is the number of work items that will be processed, not the number of worker threads.
The count_down() method could be called zero times, one time, or multiple times on each thread, and that number could be different for different threads. For example, even if you push 7 messages onto 7 threads, it might be that all 7 items are processed onto the same one thread (rather than one for each thread) and that's fine.
Other unrelated work items could be interleaved with these ones (e.g. because they weree pushed onto the queue by other producer threads) and again that's fine.
In principle, it's possible that latch.wait() won't be called until after all of the worker threads have already finished processing all of the work items. (This is the sort of odd condition you need to look out for when writing threaded code.) But that's OK, it's not a race condition: latch.wait() will just immediately return in that case.
An alternative to using a latch is that there's another queue, in addition to the one shown here, that contains the result of the work items. The thread pool callback pushes results on to that queue while the producer thread pops results off of it. Basically, it goes in the opposite direction to the queue in this code. That's a perfectly valid strategy too, in fact if anything it's more common, but there are other situations where the latch is more useful.
Barrier
A barrier is often used to make all threads wait simultaneously so that the data associated with all of the threads can be operated on simultaneously.
typedef Fn std::function<void()>;
Fn completionFn = [&procs]() {
// Do something with the whole vector of Proc objects
};
auto barrier = std::make_shared<std::barrier<Fn>>(worker_count, completionFn);
auto workerFn = [barrier](Proc&) {
barrier->count_down_and_wait();
};
for (size_t i = 0; i < worker_count; ++i) {
queue.push_back(workerFn);
}
How this works:
All of the worker threads will pop one of these workerFn items off of the queue and call barrier.count_down_and_wait().
Once all of them are waiting, one of them will call completionFn() while the others continue to wait.
Once that function completes they will all return from count_down_and_wait() and be free to pop other, unrelated, work items from the queue.
Notes:
Here the barrier count is the number of worker threads.
It is guaranteed that each thread will pop precisely one workerFn off of the queue and handle it. Once a thread has popped one off of the queue, it will wait in barrier.count_down_and_wait() until all the other copies of workerFn have been popped off by other threads, so there is no chance of it popping another one off.
I used a shared pointer to the barrier so that it will be destroyed automatically once all the work items are done. This wasn't an issue with the latch because there we could just make it a local variable in the producer thread function, because it waits until the worker threads have used the latch (it calls latch.wait()). Here the producer thread doesn't wait for the barrier so we need to manage the memory in a different way.
If you did want the original producer thread to wait until the barrier has been finished, that's fine, it can call count_down_and_wait() too, but you will obviously need to pass worker_count + 1 to the barrier's constructor. (And then you wouldn't need to use a shared pointer for the barrier.)
If other work items are being pushed onto the queue at the same time, that's fine too, although it will potentially waste time as some threads will just be sitting there waiting for the barrier to be acquired while other threads are distracted by other work before they acquire the barrier.
!!! DANGER !!!
The last bullet point about other working being pushed onto the queue being "fine" is only the case if that other work doesn't also use a barrier! If you have two different producer threads putting work items with a barrier on to the same queue and those items are interleaved, then some threads will wait on one barrier and others on the other one, and neither will ever reach the required wait count - DEADLOCK. One way to avoid this is to only ever use barriers like this from a single thread, or even to only ever use one barrier in your whole program (this sounds extreme but is actually quite a common strategy, as barriers are often used for one-time initialisation on startup). Another option, if the thread queue you're using supports it, is to atomically push all work items for the barrier onto the queue at once so they're never interleaved with any other work items. (This won't work with the moodycamel queue, which supports pushing multiple items at once but doesn't guarantee that they won't be interleved with items pushed on by other threads.)
Barrier without completion function
At the point when you asked this question, the proposed experimental API didn't support completion functions. Even the current API at least allows not using them, so I thought I should show an example of how barriers can be used like that too.
auto barrier = std::make_shared<std::barrier<>>(worker_count);
auto workerMainFn = [&procs, barrier](Proc&) {
barrier->count_down_and_wait();
// Do something with the whole vector of Proc objects
barrier->count_down_and_wait();
};
auto workerOtherFn = [barrier](Proc&) {
barrier->count_down_and_wait(); // Wait for work to start
barrier->count_down_and_wait(); // Wait for work to finish
}
queue.push_back(std::move(workerMainFn));
for (size_t i = 0; i < worker_count - 1; ++i) {
queue.push_back(workerOtherFn);
}
How this works:
The key idea is to wait for the barrier twice in each thread, and do the work in between. The first waits have the same purpose as the previous example: they ensure any earlier work items in the queue are finished before starting this work. The second waits ensure that any later items in the queue don't start until this work has finished.
Notes:
The notes are mostly the same as the previous barrier example, but here are some differences:
One difference is that, because the barrier is not tied to the specific completion function, it's more likely that you can share it between multiple uses, like we did in the latch example, avoiding the use of a shared pointer.
This example makes it look like using a barrier without a completion function is much more fiddly, but that's just because this situation isn't well suited to them. Sometimes, all you need is to reach the barrier. For example, whereas we initialised a queue before the threads started, maybe you have a queue for each thread but initialised in the threads' run functions. In that case, maybe the barrier just signifies that the queues have been initialised and are ready for other threads to pass messages to each other. In that case, you can use a barrier with no completion function without needing to wait on it twice like this.
You could actually use a latch for this, calling count_down() and then wait() in place of count_down_and_wait(). But using a barrier makes more sense, both because calling the combined function is a little simpler and because using a barrier communicates your intention better to future readers of the code.
Any any case, the "DANGER" warning from before still applies.
I have a for loop that I would like to make parallel, however the threads must share an unordered_map and a vector.
Because the for loop is somewhat big I will post here a concise overview of it so that I can make my main problem clear. Please read the comments.
unordered_map<string, vector<int>> sharedUM;
/*
here I call a function that updates the unordered_map with some
initial data, however the unordered_map will need to be updated by
the threads inside the for loop
*/
vector<int> sharedVector;
/*
the shared vector initially is empty, the threads will
fill it with integers, the order of these integers should be in ascending
order, however I can simply sort the array after all the
threads finish executing so I guess we can assume that the order
does not matter
*/
#pragma omp parallel for
for(int i=0; i<N; i++){
key = generate_a_key_value_according_to_an_algorithm();
std::unordered_map<string, vector<int>::iterator it = sharedUM.find(key);
/*
according to the data inside it->second(the value),
the thread makes some conclusions which then
uses in order to figure out whether
it should run a high complexity algorithm
or not.
*/
bool conclusion = make_conclusion();
if(conclusion == true){
results = run_expensive_algorithm();
/*
According to the results,
the thread updates some values of
the key that it previously searched for inside the unordered_map
this update may help other threads avoid running
the expensive algorithm
*/
}
sharedVector.push_back(i);
}
Initially I left the code as it is, so I just used that #pragma over the for loop, however I got a few problems regarding the update of the sharedVector. So I decided to use simple locks in order to force a thread acquire the lock before writing to the vector. So in my implementation I had something like this:
omp_lock_t sharedVectorLock;
omp_init_lock(&sharedVectorLock);
...
for(...)
...
omp_set_lock(&sharedVectorLock);
sharedVector.push_back(i);
omp_unset_lock(&sharedVectorLock);
...
omp_destroy_lock(&sharedVectorLock);
I had run my application many times and everything seemed to be working great, and that's until I decided to rerun it automatically too many times until I got wrong results. Because I'm very new to the world of OpenMP and the threads in general, I wasn't aware of the fact that we should lock all the readers when a writer is updating some shared data. As you can see here in my application the threads always read some data from the unordered_map in order make some conclusions and learn things about the key that was assigned to them. What happens though if two threads have to work with the same key, and while some other thread is trying to read the values of this key, another one has reached the point of updating those values? I believe that's where my problem occurs.
However my main problem right now is that I'm not sure what would be the best way to avoid such things from happening. It's like my system works for 99% of the time, but that 1% ruins everything because two threads are rarely assigned with the same key which in turn is because my unordered_map is usually big.
Would locking the unordered_map do my job? Most likely, but that wouldn't be efficient because a thread A that wants to work with the key x would have to wait for a thread B that is already working with the key y where y can be different than x to finish.
So my main question is, how should I approach this problem? How can I lock the unordered_map if and only if two threads are working with the same key?
Thank you in advance
1 on using locks and mutexes. You must declare and initialise the lock variables outside of the parallel block (before #pragma omp parallel) and then use them inside the parallel block: (1) acquire a lock (this may block if another thread has locked it), (2) change the variable with the race condition, (3) release the lock. Finally, destroy it after exiting the parallel block. A lock declared inside the parallel block is local to the thread and hence cannot provide synchronisation.
This may explain your problems.
2 on writing into complicated C++ containers. OpenMP was designed originally for simple FORTRAN do loops (similar to C/C++ for loops with integer control variables). Everything more complicated will give you headache. To be on the safe side, any non-constant operation on a C++ container must be performed within a lock (use the same lock for any such operation on the same container) or omp critical region (use the same name for any such operation on the same container). This includes pop() and push() etc, anything but simple reads. This can only remain efficient if such non-constant container operations take only a tiny fraction of the time.
3 If I were you, I wouldn't bother with openMP (I have used it but am regretting this now). With C++ you could use TBB, which also comes with some threadsafe but lock-free containers. It also allows you to think in terms of tasks, not threads, which are executed recursively (a parent task spawns child tasks, etc), but TBB has some simple implementations for parallel for loops, for instance.
An alternative approach would be to use TBB's concurrent_unordered_map.
You don't have to use the rest of TBB's parallelism support (though if you're starting from scratch in C++ it's certainly more "c++-ish" than OpenMP).
May be this could help:
vector<bool> sv(N);
replace
sharedVector.push_back(i);
by
sv[i]=true;
this allows to avoid locks (very time consuming) and sharedVector
can easily be sorted, e.g
for(int i=0; i<N;i++){
if(sv[i])sharedVector.push_back(i);
}
I'm fairly new to C++ standard library and have been using standard library lists for specific multithreaded implementation. I noticed that there might be a trick with using lists that I have not seen on any tutorial/blog/forum posts and though seems obvious to me does not seem to be considered by anyone. So maybe I'm too new and possibly missing something so hopefully someone smarter than me can perhaps validate what I am trying to achieve or explain to me what I am doing wrong.
So we know that in general standard library containers are not thread safe - but this seems like a guiding statement more than a rule. With lists it seems that there is a level of tolerance for thread safety. Let me explain, we know that lists do not get invalidated if we add/delete from the list. The only iterator that gets invalidated is the deleted item - which you can fix with the following line of code:
it = myList.erase(it)
So now lets say we have two threads and call them thread 1 and thread 2.
Thread 1's responsibility is to add to the list. It treats it as a queue, so it uses the std::list::push_back() function call.
Thread 2's responsibility is to process the data stored in the list as a queue and then after processing it will remove elements from the list.
Its guaranteed that Thread 2 will not remove elements in the list that were just added during its processing and Thread 1 guarantees that it will queue up the necessary data well ahead for Thread 2's processing. However, keep in mind elements can be added during Thread 2's processing.
So it seems that this is a reasonable use of lists in this multithreaded environment without the use of a locks for data protection. The reason why I say its reasonable is because, essentially, Thread 2 will only process data up to now such that it can retreive the current end iterator shown by the following pseudocode:
Thread 2 {
iter = myList.begin();
lock();
iterEnd = myList.end(); // lock data temporarily in order to get the current
// last element in the list
unlock();
// perform necessary processing
while (iter != iterEnd) {
// process data
// ...
// remove element
iter = myList.erase(iter);
}
}
Thread 2 uses a lock for a very short amount of time just to know where to stop processing, but for the most part Thread 1 and Thread 2 don't require any other locking. In addition, Thread 2 can possibly avoid locking too if its scope to know the current last element is flexible.
Does anyone see anything wrong with my suggestion?
Thanks!
Your program is racy. As an example of one obvious data race: std::list is more than just a collection of doubly-linked nodes. It also has, for example, a data member that stores the number of nodes in the list (it needs not be a single data member, but it has to store the count somewhere).
Both of your threads will modify this data member concurrently. Because there is no synchronization of those modifications, your program is racy.
Instances of the Standard Library containers cannot be mutated from multiple threads concurrently without external synchronization.
I have a custom thread pool class, that creates some threads that each wait on their own event (signal). When a new job is added to the thread pool, it wakes the first free thread so that it executes the job.
The problem is the following : I have around 1000 loops of each around 10'000 iterations do to. These loops must be executed sequentially, but I have 4 CPUs available. What I try to do is to split the 10'000 iteration loops into 4 2'500 iterations loops, ie one per thread. But I have to wait for the 4 small loops to finish before going to the next "big" iteration. This means that I can't bundle the jobs.
My problem is that using the thread pool and 4 threads is much slower than doing the jobs sequentially (having one loop executed by a separate thread is much slower than executing it directly in the main thread sequentially).
I'm on Windows, so I create events with CreateEvent() and then wait on one of them using WaitForMultipleObjects(2, handles, false, INFINITE) until the main thread calls SetEvent().
It appears that this whole event thing (along with the synchronization between the threads using critical sections) is pretty expensive !
My question is : is it normal that using events takes "a lot of" time ? If so, is there another mechanism that I could use and that would be less time-expensive ?
Here is some code to illustrate (some relevant parts copied from my thread pool class) :
// thread function
unsigned __stdcall ThreadPool::threadFunction(void* params) {
// some housekeeping
HANDLE signals[2];
signals[0] = waitSignal;
signals[1] = endSignal;
do {
// wait for one of the signals
waitResult = WaitForMultipleObjects(2, signals, false, INFINITE);
// try to get the next job parameters;
if (tp->getNextJob(threadId, data)) {
// execute job
void* output = jobFunc(data.params);
// tell thread pool that we're done and collect output
tp->collectOutput(data.ID, output);
}
tp->threadDone(threadId);
}
while (waitResult - WAIT_OBJECT_0 == 0);
// if we reach this point, endSignal was sent, so we are done !
return 0;
}
// create all threads
for (int i = 0; i < nbThreads; ++i) {
threadData data;
unsigned int threadId = 0;
char eventName[20];
sprintf_s(eventName, 20, "WaitSignal_%d", i);
data.handle = (HANDLE) _beginthreadex(NULL, 0, ThreadPool::threadFunction,
this, CREATE_SUSPENDED, &threadId);
data.threadId = threadId;
data.busy = false;
data.waitSignal = CreateEvent(NULL, true, false, eventName);
this->threads[threadId] = data;
// start thread
ResumeThread(data.handle);
}
// add job
void ThreadPool::addJob(int jobId, void* params) {
// housekeeping
EnterCriticalSection(&(this->mutex));
// first, insert parameters in the list
this->jobs.push_back(job);
// then, find the first free thread and wake it
for (it = this->threads.begin(); it != this->threads.end(); ++it) {
thread = (threadData) it->second;
if (!thread.busy) {
this->threads[thread.threadId].busy = true;
++(this->nbActiveThreads);
// wake thread such that it gets the next params and runs them
SetEvent(thread.waitSignal);
break;
}
}
LeaveCriticalSection(&(this->mutex));
}
This looks to me as a producer consumer pattern, which can be implented with two semaphores, one guarding the queue overflow, the other the empty queue.
You can find some details here.
Yes, WaitForMultipleObjects is pretty expensive. If your jobs are small, the synchronization overhead will start to overwhelm the cost of actually doing the job, as you're seeing.
One way to fix this is bundle multiple jobs into one: if you get a "small" job (however you evaluate such things), store it someplace until you have enough small jobs together to make one reasonably-sized job. Then send all of them to a worker thread for processing.
Alternately, instead of using signaling you could use a multiple-reader single-writer queue to store your jobs. In this model, each worker thread tries to grab jobs off the queue. When it finds one, it does the job; if it doesn't, it sleeps for a short period, then wakes up and tries again. This will lower your per-task overhead, but your threads will take up CPU even when there's no work to be done. It all depends on the exact nature of the problem.
Watch out, you are still asking for a next job after the endSignal is emitted.
for( ;; ) {
// wait for one of the signals
waitResult = WaitForMultipleObjects(2, signals, false, INFINITE);
if( waitResult - WAIT_OBJECT_0 != 0 )
return;
//....
}
Since you say that it is much slower in parallel than sequential execution, I assume that your processing time for your internal 2500 loop iterations is tiny (in the few micro seconds range). Then there is not much you can do except review your algorithm to split larger chunks of precessing; OpenMP won't help and every other synchronization techniques won't help either because they fundamentally all rely on events (spin loops do not qualify).
On the other hand, if your processing time of the 2500 loop iterations is larger than 100 micro seconds (on current PCs), you might be running into limitations of the hardware. If your processing uses a lot of memory bandwidth, splitting it to four processors will not give you more bandwidth, it will actually give you less because of collisions. You could also be running into problems of cache cycling where each of your top 1000 iteration will flush and reload the cache of the 4 cores. Then there is no one solution, and depending on your target hardware, there may be none.
If you are just parallelizing loops and using vs 2008, I'd suggest looking at OpenMP. If you're using visual studio 2010 beta 1, I'd suggesting looking at the parallel pattern library, particularly the "parallel for" / "parallel for each"
apis or the "task group class because these will likely do what you're attempting to do, only with less code.
Regarding your question about performance, here it really depends. You'll need to look at how much work you're scheduling during your iterations and what the costs are. WaitForMultipleObjects can be quite expensive if you hit it a lot and your work is small which is why I suggest using an implementation already built. You also need to ensure that you aren't running in debug mode, under a debugger and that the tasks themselves aren't blocking on a lock, I/O or memory allocation, and you aren't hitting false sharing. Each of these has the potential to destroy scalability.
I'd suggest looking at this under a profiler like xperf the f1 profiler in visual studio 2010 beta 1 (it has 2 new concurrency modes which help see contention) or Intel's vtune.
You could also share the code that you're running in the tasks, so folks could get a better idea of what you're doing, because the answer I always get with performance issues is first "it depends" and second, "have you profiled it."
Good Luck
-Rick
It shouldn't be that expensive, but if your job takes hardly any time at all, then the overhead of the threads and sync objects will become significant. Thread pools like this work much better for longer-processing jobs or for those that use a lot of IO instead of CPU resources. If you are CPU-bound when processing a job, ensure you only have 1 thread per CPU.
There may be other issues, how does getNextJob get its data to process? If there's a large amount of data copying, then you've increased your overhead significantly again.
I would optimise it by letting each thread keep pulling jobs off the queue until the queue is empty. that way, you can pass a hundred jobs to the thread pool and the sync objects will be used just the once to kick off the thread. I'd also store the jobs in a queue and pass a pointer, reference or iterator to them to the thread instead of copying the data.
The context switching between threads can be expensive too. It is interesting in some cases to develop a framework you can use to process your jobs sequentially with one thread or with multiple threads. This way you can have the best of the two worlds.
By the way, what is your question exactly ? I will be able to answer more precisely with a more precise question :)
EDIT:
The events part can consume more than your processing in some cases, but should not be that expensive, unless your processing is really fast to achieve. In this case, switching between thredas is expensive too, hence my answer first part on doing things sequencially ...
You should look for inter-threads synchronisation bottlenecks. You can trace threads waiting times to begin with ...
EDIT: After more hints ...
If I guess correctly, your problem is to efficiently use all your computer cores/processors to parralellize some processing essencialy sequential.
Take that your have 4 cores and 10000 loops to compute as in your example (in a comment). You said that you need to wait for the 4 threads to end before going on. Then you can simplify your synchronisation process. You just need to give your four threads thr nth, nth+1, nth+2, nth+3 loops, wait for the four threads to complete then going on. You should use a rendezvous or barrier (a synchronization mechanism that wait for n threads to complete). Boost has such a mechanism. You can look the windows implementation for efficiency. Your thread pool is not really suited to the task. The search for an available thread in a critical section is what is killing your CPU time. Not the event part.
It appears that this whole event thing
(along with the synchronization
between the threads using critical
sections) is pretty expensive !
"Expensive" is a relative term. Are jets expensive? Are cars? or bicycles... shoes...?
In this case, the question is: are events "expensive" relative to the time taken for JobFunction to execute? It would help to publish some absolute figures: How long does the process take when "unthreaded"? Is it months, or a few femtoseconds?
What happens to the time as you increase the threadpool size? Try a pool size of 1, then 2 then 4, etc.
Also, as you've had some issues with threadpools here in the past, I'd suggest some debug
to count the number of times that your threadfunction is actually invoked... does it match what you expect?
Picking a figure out of the air (without knowing anything about your target system, and assuming you're not doing anything 'huge' in code you haven't shown), I'd expect the "event overhead" of each "job" to be measured in microseconds. Maybe a hundred or so. If the time taken to perform the algorithm in JobFunction is not significantly MORE than this time, then your threads are likely to cost you time rather than save it.
As mentioned previously, the amount of overhead added by threading depends on the relative amount of time taken to do the "jobs" that you defined. So it is important to find a balance in the size of the work chunks that minimizes the number of pieces but does not leave processors idle waiting for the last group of computations to complete.
Your coding approach has increased the amount of overhead work by actively looking for an idle thread to supply with new work. The operating system is already keeping track of that and doing it a lot more efficiently. Also, your function ThreadPool::addJob() may find that all of the threads are in use and be unable to delegate the work. But it does not provide any return code related to that issue. If you are not checking for this condition in some way and are not noticing errors in the results, it means that there are idle processors always. I would suggest reorganizing the code so that addJob() does what it is named -- adds a job ONLY (without finding or even caring who does the job) while each worker thread actively gets new work when it is done with its existing work.