I'm building a simple C++ server where I receive an image via a POST request, send it to a GPU for processing, and once I have the results from the GPU, send back a response.
To be able to handle many simultaneous connections (and to learn something new), I'm using Boost::asio, following the server4 example (link) that uses boost::asio::coroutine. The issue I'm running into is how to send data from the coroutine to the GPU while not blocking the coroutine.
The GPU is most efficient when it can process a batch of requests together. But even processing one-request-at-a-time (cannot be context-switched as a CPU and memory IO is the bottleneck), it must process a complete request before starting on the next one. This means I need to queue the requests from the coroutines, and signal the coroutine somehow when the GPU processing is complete.
Been looking through the boost::asio reference but nothing is jumping out at me. In general, how are boost asio coroutines used where a response cannot be generated immediately (e.g. may take ~500ms). How is this typically done?
Set up the part of the program that interacts with the GPU to run on it's own thread. When you recieve a request that requires the GPU, box up the image data, along with the socket object and other required parameters, and push it into a (probably mutex-locked) queue. The gpu thread can then pull from the queue, perform its work, then spawn a new task on the io_service to finish the transaction with the client, using the information stored in the structure.
The gpu thread can be set to sleep until work becomes available using a condition variable and having the network code invoke notify_one() after it has pushed the work request.
Another trick that works exactly as you describe is to set a deadline timer with an infinite deadline, which you await in the coro (yielding).
Then from the GPU task you cancel the time when the work is complete, resuming the coro.
See for example
Why does Boost.Asio not support an Event-based interface (for example, something along the lines of condition variable signaling)?
Can a boost::asio::yield_context be used as a deadline_timer handler when doing cancel?
Related
I'm creating an async gRPC server in C++. One of the methods streams data from the server to clients - it's used to send data updates to clients. The frequency of the data updates isn't predictable. They could be nearly continuous or as infrequent as once per hour. The model used in the gRPC example with the "CallData" class and the CREATE/PROCESS/FINISH states doesn't seem like it would work very well for that. I've seen an example that shows how to create a 'polling' loop that sleeps for some time and then wakes up to check for new data, but that doesn't seem very efficient.
Is there another way to do this? If I use the "CallData" method can it block in the 'PROCESS' state until there's data (which probably wouldn't be my first choice)? Or better, can I structure my code so I can notify a gRPC handler when data is available?
Any ideas or examples would be appreciated.
In a server-side streaming example, you probably need more states, because you need to track whether there is currently a write already in progress. I would add two states, one called WRITE_PENDING that is used when a write is in progress, and another called WRITABLE that is used when a new message can be sent immediately. When a new message is produced, if you are in state WRITABLE, you can send immediately and go into state WRITE_PENDING, but if you are in state WRITE_PENDING, then the newly produced message needs to go into a queue to be sent after the current write finishes. When a write finishes, if the queue is non-empty, you can grab the next message from the queue and immediately start a write for it; otherwise, you can just go into state WRITABLE and wait for another message to be produced.
There should be no need to block here, and you probably don't want to do that anyway, because it would tie up a thread that should otherwise be polling the completion queue. If all of your threads wind up blocked that way, you will be blind to new events (such as new calls coming in).
An alternative here would be to use the C++ sync API, which is much easier to use. In that case, you can simply write straight-line blocking code. But the cost is that it creates one thread on the server for each in-progress call, so it may not be feasible, depending on the amount of traffic you're handling.
I hope this information is helpful!
I have a single threaded asynchronous tcp server written using boost asio. Each incoming request will go through several processing steps (synchronous and asynchronous) and finally send back the response using async write.
For small loads with 10 concurrent requests, it works decently. However, when I test using a parallelism of 100, things start worsening. Response latency starts increasing as time progresses. So, I want to try with some multi-threaded processing for handling requests.
I am looking for a decent example / help on creating and running multiple threads for asynchronous reading/writing to clients. I have the following doubts:
Should I use a single IOS object and call its run method in all of the threads of the thread pool, or should I use a separate IOS per thread?
If I use a single IOS, is there a possibility that part of the tcp data goes to one thread, while another part going to another thread and so on.. Is this understanding correct?
Is there any other better way?
Thanks for any help and pointers here.
Without seeing your code I can only guess what goes wrong. Most probably you're running long actions inside async completion handlers. The completion handlers should be fast - get the data, hand it off for further processing, done.
As a first priority, I would go full-asynchronous and run all processing in a thread pool. You can find an example here, where a new thread is started for every new client, which you can replace with a thread pool.
Use a single io_service. A single io_service can handle a lot of parallelism, provided you don't delay it inside completion handlers. This simplifies the implementation because you don't have to worry about completion handlers running in parallel, which will happen if you run multiple IOS in multiple threads.
Q1: Should I use a single IOS object and call its run method in all of the threads of the thread pool, or should I use a separate IOS per thread?
Either you can
HTTP Server 2 - IOS per thread
HTTP Server 3 - single IOS with thread pool
Q2: If I use a single IOS, is there a possibility that part of the tcp data goes to one thread, while another part going to another thread and so on.. Is this understanding correct?
Yes, there is a race condition, but boost.asio support strand to avoid it.
Q3: Is there any other better way?
To me, not find a better way, if you find, tell me or past here, thank you.
BTW, as #rustyx said, your program is blocked at sync calls, turn to full-asynchronous calls will help.
I am working on a multithreaded middleware enviornment. The framework is basically a capturing and streaming framework. So it involves a number of threads.
To give you all a brief idea of the threading architecture:
There are seprate threads for demultiplexer, receiveVideo, DecodeVideo, DisplayVideo etc. Each thread performs its functionlity, for eg:
demultiplexer extracts audio, video packets
receivevideo receives header + payload of video packet & removes payload
DecodeVideo receives payload & decodes payload packet
DisplayVideo receives decoded packets & displays the decoded packets on display
Thus each thread feeds the extracted data to the next thread. The threads share data buffers amongst them and the buffers are synchronised through use of mutexes and semaphores. Similarly, there are other threads for handling ananlogvideo and analogaudio etc.
All the threads are spawned in during initialization but they remain blocked on a semaphore and depending upon the input(analog/digitial) selective semaphores are signalled so that specifc threads get unblocked & move on to do their work. At various stages each thread calls some lower level(driver calls)to get data or write data etc. These calls are blocking and the errors resulting from these calls(driver returning corrupted data, driver stalling) should be handled but are not being handled currently.
I wanted to implement a thread monitoring mechanism where a thread will monitor these worker threads and if an error condition occurs will take some preventive actions. As I understand certain such mechanisms are commonly used like Watchdogs in UI or MMI applications. I am trying to look for something similar.
I am using pthreads and No Boost or STL(its a legacy code, pretty much procedural C++)
Any ideas about specific framework or design patterns or open source projects which do something similar and might help in with ideas for implementing my requirement?
Can you ping the threads - periodically send each one a message on its usual input queue, interleaved with all the other normal stuff, asking it to return its status? When each handler thread gets the message, it loads the message with status stuff - how many messages its processed since the last ping, length of its input/output queue, last time that its driver returned OK, that sort of stats - and queues it back to your Thread Monitoring Mechanism. Your TMM would have to time out the replies in case some thread/s is/are stuck.
You could, maybe, just post one message down the whole chain, each thread adding its own status in different fields. That would mean only one timeout, after which your TMM would have to examine the message to see how far down the chain it got.
There are other things - I like to keep an on-screen dump, on a 1s timer, of the length of queues and depth of buffer pools. If something stuffs, I can usually tell roughly where it is, (eg. a pool is emptying and some queue is growing - the queue comsumer is wasted).
Rgds,
Martin
What about using a signalling system to wake up your monitoring thread when something's gone awry in one of your worker threads. You can emulate the signalling with an ResetEvent of some type.
When an exception occurs in your worker thread, you have some data structure you fill up with the data about the exception and then you can pass that on to your monitoring thread. You wake up the monitoring thread by using the event.
Then the monitoring thread can do what you need it to do.
I'm guessing you don't wish to have your monitoring thread active unless something has gone wrong, right?
I would like to spawn off threads to perform certain tasks, and use a thread-safe queue to communicate with them. I would also like to be doing IO to a variety of file descriptors while I'm waiting.
What's the recommended way to accomplish this? Do I have to created an inter-thread pipe and write to it when the queue goes from no elements to some elements? Isn't there a better way?
And if I have to create the inter-thread pipe, why don't more libraries that implement shared queues allow you to create the shared queue and inter-thread pipe as a single entity?
Does the fact I want to do this at all imply a fundamental design flaw?
I'm asking this about both C++ and Python. And I'm mildly interested in a cross-platform solution, but primarily interested in Linux.
For a more concrete example...
I have some code which will be searching for stuff in a filesystem tree. I have several communications channels open to the outside world through sockets. Requests that may (or may not) result in a need to search for stuff in the filesystem tree will be arriving.
I'm going to isolate the code that searches for stuff in the filesystem tree in one or more threads. I would like to take requests that result in a need to search the tree and put them in a thread-safe queue of things to be done by the searcher threads. The results will be put into a queue of completed searches.
I would like to be able to service all the non-search requests quickly while the searches are going on. I would like to be able to act on the search results in a timely fashion.
Servicing the incoming requests would generally imply some kind of event-driven architecture that uses epoll. The queue of disk-search requests and the return queue of results would imply a thread-safe queue that uses mutexes or semaphores to implement the thread safety.
The standard way to wait on an empty queue is to use a condition variable. But that won't work if I need to service other requests while I'm waiting. Either I end up polling the results queue all the time (and delaying the results by half the poll interval, on average), blocking and not servicing requests.
Whenever one uses an event driven architecture, one is required to have a single mechanism to report event completion. On Linux, if one is using files, one is required to use something from the select or poll family meaning that one is stuck with using a pipe to initiate all none file related events.
Edit: Linux has eventfd and timerfd. These can be added to your epoll list and used to break out of the epoll_wait when either triggered from another thread or on a timer event respectively.
There is another option and that is signals. One can use fcntl modify the file descriptor such that a signal is emitted when the file descriptor becomes active. The signal handler may then push a file-ready message onto any type of queue of your choosing. This may be a simple semaphore or mutex/condvar driven queue. Since one is now no longer using select/poll, one no longer needs to use a pipe to queue none file based messages.
Health warning: I have not tried this and although I cannot see why it will not work, I don't really know the performance implications of the signal approach.
Edit: Manipulating a mutex in a signal handler is probably a very bad idea.
I've solved this exact problem using what you mention, pipe() and libevent (which wraps epoll). The worker thread writes a byte to its pipe FD when its output queue goes from empty to non-empty. That wakes up the main IO thread, which can then grab the worker thread's output. This works great is actually very simple to code.
You have the Linux tag so I am going to throw this out: POSIX Message Queues do all this, which should fulfill your "built-in" request if not your less desired cross-platform wish.
The thread-safe synchronization is built-in. You can have your worker threads block on read of the queue. Alternatively MQs can use mq_notify() to spawn a new thread (or signal an existing one) when there is a new item put in the queue. And since it looks like you are going to be using select(), MQ's identifier (mqd_t) can be used as a file descriptor with select.
It seems nobody has mentioned this option yet:
Don't run select/poll/etc. in your "main thread". Start a dedicated secondary thread which does the I/O and pushes notifications into your thread-safe queue (the same queue which your other threads use to communicate with the main thread) when I/O operations complete.
Then your main thread just needs to wait on the notification queue.
Duck's and twk's are actually better answers than doron's (the one selected by the OP), in my opinion. doron suggests writing to a message queue from within the context of a signal handler, and states that the message queue can be "any type of queue." I would strongly caution you against this since many C library/system calls cannot safely be called from within a signal handler (see async-signal-safe).
In particuliar, if you choose a queue protected by a mutex, you should not access it from a signal handler. Consider this scenario: your consumer thread locks the queue to read it. Immediately after, the kernel delivers the signal to notify you that a file descriptor now has data on it. You signal handler runs in the consumer thread, necessarily), and tries to put something on your queue. To do this, it first has to take the lock. But it already holds the lock, so you are now deadlocked.
select/poll is, in my experience, the only viable solution to an event-driven program in UNIX/Linux. I wish there were a better way inside a mutlithreaded program, but you need some mechanism to "wake up" your consumer thread. I have yet to find a method that does not involve a system call (since the consumer thread is on a waitqueue inside the kernel during any blocking call such as select).
EDIT: I forgot to mention one Linux-specific way to handle signals when using select/poll: signalfd(2). You get a file descriptor you can select/poll on, and you handling code runs normally instead of in a signal handler's context.
This is a very common seen problem, especially when you are developing network server-side program. Most Linux server-side program's main look will loop like this:
epoll_add(serv_sock);
while(1){
ret = epoll_wait();
foreach(ret as fd){
req = fd.read();
resp = proc(req);
fd.send(resp);
}
}
It is single threaded(the main thread), epoll based server framework. The problem is, it is single threaded, not multi-threaded. It requires that proc() should never blocks or runs for a significant time(say 10 ms for common cases).
If proc() will ever runs for a long time, WE NEED MULTI THREADS, and executes proc() in a separated thread(the worker thread).
We can submit task to the worker thread without blocking the main thread, using a mutex based message queue, it is fast enough.
epoll_add(serv_sock);
while(1){
ret = epoll_wait();
foreach(ret as fd){
req = fd.read();
queue.add_job(req); // fast, non blockable
}
}
Then we need a way to obtain the task result from a worker thread. How? If we just check the message queue directly, before or after epoll_wait().
epoll_add(serv_sock);
while(1){
ret = epoll_wait(); // may blocks for 10ms
resp = queue.check_result(); // fast, non blockable
foreach(ret as fd){
req = fd.read();
queue.add_job(req); // fast, non blockable
}
}
However, the checking action will execute after epoll_wait() to end, and epoll_wait() usually blocks for 10 micro seconds(common cases) if all file descriptors it waits are not active.
For a server, 10 ms is quite a long time! Can we signal epoll_wait() to end immediately when task result is generated?
Yes! I will describe how it is done in one of my open source project:
Create a pipe for all worker threads, and epoll waits on that pipe as well. Once a task result is generated, the worker thread writes one byte into the pipe, then epoll_wait() will end in nearly the same time! - Linux pipe has 5 us to 20 us latency.
In my project SSDB(a Redis protocol compatible in-disk NoSQL database), I create a SelectableQueue for passing messages between the main thread and worker threads. Just like its name, SelectableQueue has an file descriptor, which can be wait by epoll.
SelectableQueue: https://github.com/ideawu/ssdb/blob/master/src/util/thread.h#L94
Usage in main thread:
epoll_add(serv_sock);
epoll_add(queue->fd());
while(1){
ret = epoll_wait();
foreach(ret as fd){
if(fd is queue){
sock, resp = queue->pop_result();
sock.send(resp);
}
if(fd is client_socket){
req = fd.read();
queue->add_task(fd, req);
}
}
}
Usage in worker thread:
fd, req = queue->pop_task();
resp = proc(req);
queue->add_result(fd, resp);
C++11 has std::mutex and std::condition_variable. The two can be used to have one thread signal another when a certain condition is met. It sounds to me like you will need to build your solution out of these primitives. If you environment does not yet support these C++11 library features, you can find very similar ones at boost. Sorry, can't say much about python.
One way to accomplish what you're looking to do is by implementing the Observer Pattern
You would register your main thread as an observer with all your spawned threads, and have them notify it when they were done doing what they were supposed to (or updating during their run with the info you need).
Basically, you want to change your approach to an event-driven model.
I'm programming an online game for two reasons, one to familiarize myself with server/client requests in a realtime environment (as opposed to something like a typical web browser, which is not realtime) and to actually get my hands wet in that area, so I can proceed to actually properly design one.
Anywho, I'm doing this in C++, and I've been using winsock to handle my basic, basic network tests. I obviously want to use a framelimiter and have 3D going and all of that at some point, and my main issue is that when I do a send() or receive(), the program kindly idles there and waits for a response. That would lead to maybe 8 fps on even the best internet connection.
So the obvious solution to me is to take the networking code out of the main process and start it up in its own thread. Ideally, I would call a "send" in my main process which would pass the networking thread a pointer to the message, and then periodically (every frame) check to see if the networking thread had received the reply, or timed out, or what have you. In a perfect world, I would actually have 2 or more networking threads running simultaneously, so that I could say run a chat window and do a background download of a piece of armor and still allow the player to run around all at once.
The bulk of my problem is that this is a new thing to me. I understand the concept of threading, but I can see some serious issues, like what happens if two threads try to read/write the same memory address at the same time, etc. I know that there are already methods in place to handle this sort of thing, so I'm looking for suggestions on the best way to implement something like this. Basically, I need thread A to be able to start a process in thread B by sending a chunk of data, poll thread B's status, and then receive the reply, also as a chunk of data., ideally without any major crashing going on. ^_^ I'll worry about what that data actually contains and how to handle dropped packets, etc later, I just need to get that happening first.
Thanks for any help/advice.
PS: Just thought about this, may make the question simpler. Is there a way to use the windows event handling system to my advantage? Like, would it be possible to have thread A initialize data somewhere, then trigger an event in thread B to have it pick up the data, and vice versa for thread B to tell thread A it was done? That would probably solve a lot of my problems, since I don't really need both threads to be able to work on the data at the same time, more of a baton pass really. I just don't know if this is possible between two different threads. (I know one thread can create its own messages for the event handler.)
The easiest thing
for you to do, would be to simply invoke the windows API QueueUserWorkItem. All you have to specify is the function that the thread will execute and the input passed to it. A thread pool will be automatically created for you and the jobs executed in it. New threads will be created as and when is required.
http://msdn.microsoft.com/en-us/library/ms684957(VS.85).aspx
More Control
You could have a more detailed control using another set of API's which can again manage the thread pool for you -
http://msdn.microsoft.com/en-us/library/ms686980(VS.85).aspx
Do it yourself
If you want to control all aspects of your thread creation and the pool management you would have to create the threads yourself, decide how they should end , how many to create etc (beginthreadex is the api you should be using to create threads. If you use MFC you should use AfxBeginThread function).
Send jobs to worker threads - Io completion Ports
In this case, you would also have to worry about how to communicate your jobs - i would recommend IoCOmpletionPorts to do that. It is the most scalable notification mechanism that i currently know of made for this purpose. It has the additional advantage that it is implemented in the kernel so you avoid all kinds of dead loack sitautions you would encounter if you decide to handroll something yourself.
This article will show you how with code samples -
http://blogs.msdn.com/larryosterman/archive/2004/03/29/101329.aspx
Communicate Back - Windows Messages
You could use windows messages to communicate the status back to your parent thread since it is doing the message wait anyway. use the PostMessage function to do this. (and check for errors)
ps : You could also allocate the data that needs to be sent out on a dedicated pointer and then the worker thread could take care of deleting it after sending it out. That way you avoid the return pointer traffic too.
BlodBath's suggestion of non-blocking sockets is potentially the right approach.
If you're trying to avoid using a multithreaded approach, then you could investigate the use of setting up overlapped I/O on your sockets. They will not block when you do a transmit or receive, but have the added bonus of giving you the option of waiting for multiple events within your single event loop. When your transmit has finished, you will receive an event. (see this for some details)
This is not incompatible with a multithreaded approach, so there's the option of changing your mind later. ;-)
On the design of your multithreaded app. the best thing to do is to work out all of the external activities that you want to be alerted to. For example, so far in your question you've listed network transmits, network receives, and user activity.
Depending on the number of concurrent connections you're going to be dealing with you'll probably find it conceptually simpler to have a thread per socket (assuming small numbers of sockets), where each thread is responsible for all of the processing for that socket.
Then you can implement some form of messaging system between your threads as RC suggested.
Arrange your system so that when a message is sent to a particular thread and event is also sent. Your threads can then be sent to sleep waiting for one of those events. (as well as any other stimulus - like socket events, user events etc.)
You're quite right that you need to be careful of situations where more than one thread is trying to access the same piece of memory. Mutexes and semaphores are the things to use there.
Also be aware of the limitations that your gui has when it comes to multithreading.
Some discussion on the subject can be found in this question.
But the abbreviated version is that most (and Windows is one of these) GUIs don't allow multiple threads to perform GUI operations simultaneously. To get around this problem you can make use of the message pump in your application, by sending custom messages to your gui thread to get it to perform gui operations.
I suggest looking into non-blocking sockets for the quick fix. Using non-blocking sockets send() and recv() do not block, and using the select() function you can get any waiting data every frame.
See it as a producer-consumer problem: when receiving, your network communication thread is the producer whereas the UI thread is the consumer. When sending, it's just the opposite. Implement a simple buffer class which gives you methods like push and pop (pop should be blocking for the network thread and non-blocking for the UI thread).
Rather than using the Windows event system, I would prefer something that is more portable, for example Boost condition variables.
I don't code games, but I've used a system similar to what pukku suggested. It lends nicely to doing things like having the buffer prioritize your messages to be processed if you have such a need.
I think of them as mailboxes per thread. You want to send a packet? Have the ProcessThread create a "thread message" with the payload to go on the wire and "send" it to the NetworkThread (i.e. push it on the NetworkThread's queue/mailbox and signal the condition variable of the NetworkThread so he'll wake up and pull it off). When the NetworkThread receives the response, package it up in a thread message and send it back to the ProcessThread in the same manner. Difference is the ProcessThread won't be blocked on a condition variable, just polling on mailbox.empty( ) when you want to check for the response.
You may want to push and pop directly, but a more convenient way for larger projects is to implement a toThreadName, fromThreadName scheme in a ThreadMsg base class, and a Post Office that threads register their Mailbox with. The PostOffice then has a send(ThreadMsg*); function that gets/pushes the messages to the appropriate Mailbox based on the to and from. Mailbox (the buffer/queue class) contains the ThreadMsg* = receiveMessage(), basically popping it off the underlying queue.
Depending on your needs, you could have ThreadMsg contain a virtual function process(..) that could be overridden accordingly in derived classes, or just have an ordinary ThreadMessage class with a to, from members and a getPayload( ) function to get back the raw data and deal with it directly in the ProcessThread.
Hope this helps.
Some topics you might be interested in:
mutex: A mutex allows you to lock access to specific resources for one thread only
semaphore: A way to determine how many users a certain resource still has (=how many threads are accessing it) and a way for threads to access a resource. A mutex is a special case of a semaphore.
critical section: a mutex-protected piece of code (street with only one lane) that can only be travelled by one thread at a time.
message queue: a way of distributing messages in a centralized queue
inter-process communication (IPC) - a way of threads and processes to communicate with each other through named pipes, shared memory and many other ways (it's more of a concept than a special technique)
All topics in bold print can be easily looked up on a search engine.