I've read the C10K doc as well as many related papers on scaling up a socket server. All roads point to the following:
Avoid the classic mistake of "thread per connection".
Prefer epoll over select.
Likewise, legacy async io mechanism in unix may be hard to use.
My simple TCP server just listens for client connections on a listen socket on a dedicated port. Upon receiving a new connection, parses the request, and sends a response back. Then gracefully closes the socket.
I think I have a good handle on how to scale this up on a single thread using epoll. Just one loop that calls epoll_wait for the listen socket as well as for the existing client connections. Upon return, the code will handle new creating new client connections as well as managing state of existing connections depending on which socket just got signaled. And perhaps some logic to manage connection timeouts, graceful closing of sockets, and efficient resource allocation for each connection. Seems straightforward enough.
But what if I want to scale this to take advantage of multiple threads and multiple cpu cores? The core idea that springs to mind is this:
One dedicated thread for listening for incoming connections on the TCP listen socket. Then a set of N threads (or thread pool) to handle all the active concurrent client connections. Then invent some thread safe way in which the listen thread will "dispatch" the new connection (socket) to one of the available worker threads. (ala IOCP in Windows). The worker thread will use an epoll loop on all the connections it is handling to do what the single threaded approach would do.
Am I on the right track? Or is there a standard design pattern for doing a TCP server with epoll on multiple threads?
Suggestions on how the listen thread would dispatch a new connection to the thread pool?
Firstly, note that it's C*10K*. Don't concern yourself if you're less than about 100 (on a typical system). Even then it depends on what your sockets are doing.
Yes, but keep in mind that epoll manipulation requires system calls, and their cost may or may not be more expensive than the cost of managing a few fd_sets yourself. The same goes for poll. At low counts its cheaper to be doing the processing in user space each iteration.
Asynchronous IO is very painful when you're not constrained to just a few sockets that you can juggle as required. Most people cope by using event loops, but this fragments and inverts your program flow. It also usually requires making use of large, unwieldy frameworks for this purpose since a reliable and fast event loop is not easy to get right.
The first question is, do you need this? If you're handily coping with the existing traffic by spawning off threads to handle each incoming request, then keep doing it this way. The code will be simpler for it, and all your libraries will play nicely.
As I mentioned above, juggling simultaneous requests can be complex. If you want to do this in a single loop, you'll also need to make guarantees about CPU starvation when generating your responses.
The dispatch model you proposed is the typical first step solution if your responses are expensive to generate. You can either fork or use threads. The cost of forking or generating a thread should not be a consideration in selecting a pooling mechanism: rather you should use such a mechanism to limit or order the load placed on the system.
Batching sockets onto multiple epoll loops is excessive. Use multiple processes if you're this desperate. Note that it's possible to accept on a socket from multiple threads and processes.
I would guess you are on the right track. But I also think details depend upon the particular situation (bandwidh, request patterns, indifidual request processing, etc.). I think you should try, and benchmark carefully.
Related
Now I am building an application to send big data from client to server via UDP. I have some questions are:
Should I use one thread to send data or multi-threads to send data?
If I should use multi-threads to send data, I will use one socket for all threads or one socket per one thread?
Thanks,
Should I use one thread to send data or multi-threads to send data?
Either way can work, so it's mostly a matter of personal preference. If it was me, I would use a single thread rather than multiple threads, because multiple threads are a lot harder to implement correctly, and in this case they won't buy you any additional performance, since your throughput bottleneck is almost certainly going to be either your hard disk or your network card, not the speed of your CPU core(s).
If I should use multi-threads to send data, I will use one socket for all threads or one socket per one thread?
Again, either way will work (for UDP), but if it was me, I would use one socket per thread, only because then you don't have to worry so much about race conditions during process-setup and process-shutdown (i.e. each thread simply creates and destroys its own separate/private socket, so there's no worrying about who does what to the socket when)
I am working on a project which goal is to receive and store real time data from financial exchanges, using websockets. I have some very general questions about the technology.
Suppose that I have two websocket connections open, receiving real time data from two different servers. How do I make sure not to miss any messages? I have learned a bit of asynchronous programming (python asyncio) but it does not seem to solve the problem: when I listen to one connection, I cannot listen to the other one at the same time, right?
I can think of two solutions: the first one would require that the servers use a buffer system to send their data, but I do not think this is the case (Binance, Bitfinex...). The second solution I see is to listen each websocket using a different core. If my laptop has 8 cores I can listen to 8 connections and be sure not to miss any messages. I guess I can then scale up by using a cloud service.
Is that correct or am I missing something? Many thanks.
when I listen to one connection, I cannot listen to the other one at the same time, right?
Wrong.
When using an evented programming design, you will be using an IO "reactor" that adds IO related events to the event loop.
This allows your code to react to events from a number of connections.
It's true that the code reacts to the events in sequence, but as long as your code doesn't "block", these events could be handled swiftly and efficiently.
Blocking code should be avoided and big / complicated tasks should be fragmented into a number of "events". There should be no point at which your code is "blocking" (waiting) on an IO read or write.
This will allow your code to handle all the connections without significant delays.
...the first one would require that the servers use a buffer system to send their data...
Many evented frameworks use an internal buffer that streams to the IO when "ready" events are raised. For example, look up the drained event in node.js (or the on_ready in facil.io).
This is a convenience feature rather than a requirement.
The event loop might as well add an "on ready" event and assume your code will handle buffering after partial write calls return EAGAIN / EWOULDBLOCK.
The second solution I see is to listen each websocket using a different core.
No need. A single thread on a single core with an evented design should support thousands (and tens of thousands) of concurrent clients with reasonable loads (per-client load is a significant performance factor).
Attaching TCP/IP connections to a specific core can (sometimes) improve performance, but this is a many-to-one relationship. If we had to dedicate a CPU core per connection than server prices would shoot through the roof.
I'm developing a server application using C++. I designed it in a such way that there will be main process, responsible for maintaining child processes (workers). Workers accept() new connections and create threads for handle them individually.
Suppose I create a listener socket in main process and each worker would monitor it (using kqueue, epoll, etc.) for new connections. After researching a bit, I found some affirmations of the need of using mutex on listener socket to prevent concurrent accept()s that would lead workers accept()ing the same connections at same time.
Well, being aware of such need, I'm not sure what is the best way to distribute client connections among workers, as the result will be the same as accept() them on main process and send somehow just the new socket FD to workers (new connections handling becomes blocking - one accept() at a time).
My question is: Is mutex on listening socket really needed? Am I right of its accept() blocking (one new connection accept()ed at a time) side effect?
I'm concerned about this single detail because this application must scale to up to thousands of new connections per second (exact number may vary, as this applications is intended to be used on networks with from 100s to 1000s of clients).
A long time ago there were operating systems that had race conditions if multiple processes performed an accept concurrently on the same socket. Apache used to have an optional accept mutex to resolve this.
This problem has long since been solved on every operating system you're likely to use and it's perfectly reasonable to use a shared socket that workers call accept on. If you want each worker to handle only one connection at a time, an idle worker can block in accept on a shared socket.
I'm concerned about this single detail because this application must scale to up to hundred of thousands or even millions of new connections per second. I want to avoid the work of writing two complex applications for the sole purpose of comparing both methods performance. Also, I've no way to simulate real world simultaneous connections.
You can't have it both ways. Either you abandon such ambitious scaling plans or you accept that you will have numerous major efforts on your hand. Just simulating that kind of connection load for testing would be a major effort.
I can't answer the part of your question about how threadsafe the listen() and accept() calls are, because I would never even consider trying that. What I would do is have the main thread doing the listen() and accept(), and forking a new thread when accept() returns, passing the socket off to the thread.
Similarly, you could have a bunch of running threads, and mutex a variable that will do the socket notification. Basically the same as above, but rather than create a thread at accept time, you notify an already running thread of the socket descriptor. General pseudocode might be:
main()
{
listen();
while(true)
{
int socket = accept();
if(fork() == 0)
{
DoMyThing(socket);
}
}
}
I am trying to write a server program which supports one client till now and over the few days i was trying to develop it, I concluded i needed threads. The reason for such a decision was since I take input from a wifi socket and later process it and finally write to a file, the processing time is slow and hence i needed a input thread -> circular buffer -> output thread pattern with producer consumer model which is quite common in network programming.
Now, The situation becomes complicated, as I need to manage client disconnection and re connection. I thought of using pthread_exit() and cleaning up all the semaphores and then re initializing them each time the single client re connects.
My question is that is this a efficient approach i.e. everytime killing the threads and semaphores and re creating them. Are there any better solutions.
Thanks.
My question is that is this a efficient approach i.e. everytime killing the threads and semaphores and re creating them. Are there any better solutions.
Learn how to use non-blocking sockets and an event loop. Or use a library that provides TCP sessions for you using non-blocking sockets under the hood. Such as boost::asio.
Learn how to use multi-threading without polluting your code with any synchronization primitives by using message passing to communicate between threads, not shared state. The event loop library you use for non-blocking I/O should also provide means for cross-thread message passing.
Some comments and suggestions.
1-In TCP detecting that the other side has silently disconnected it very difficult if not impossible. A client could disconnect sending a RST TCP message to the server or sending a FIN message, this is the good case. Sometimes the client can disconnect without notice (crash, cable disconnection, etc).
One suggestion here is that you consider the way client and server will communicate. For example, you can use function “select” to set a timeout for receiving a message from client and detect a silent client.
Additionally, depending on the programming language and operating system you may need to handle broken pipe (SIGPIPE) signal (in Linux, with C/C++), for a server trying to send a message through a connection closed by the client.
2-Regarding semaphores, you shouldn’t need to clean semaphores in any especial way when a client disconnect. By applying common good practices of locking and unlocking mutexes should be enough. Also with resources like file descriptors, you need to release them before ending the thread either by returning from the thread start function or with pthread_exit. Maybe I didn’t understand this part of the question.
3-Regarding threads: if you work with multiple threads to optimum is to have a pool of pre-created consumer/worker threads that will check the circular buffer to consume the next available connection. Creating and destroying threads is costly for the operating system.
Threads are resource consuming and you may exhaust operating system resources if you need to create 1,000 threads for example.
Another alternative, is to have only one consumer thread that manages all connections (sockets) asynchronously: a) Each connection has its own state. b) The main thread goes through all connections and use function “select” to detect when connection reads or a writes are ready. 3)Use of non-blocking sockets but this is not essential because from select you know which sockets are ready and will not block.
You can use functions select, poll, epoll.
One link about select and non-blocking sockets: Using select() for non-blocking sockets
Other link with an example: http://linux.die.net/man/2/select
I have created one application with server and client class which have methods for creating either creating a tcp socket or udp socket. Now my requirement is i have created two application instances of this application. Since application is in c++ in unix environment I am using putty software to run the application. I have opened two instances of putty. But now my requirement is as follows:
There can be multiple communication instances between the 2 application instances
Each communication instance, There can be multiple communication instances between the 2 application instances
Each communication instance, can be either UDP or TCP (determined from the config file)be either UDP or TCP (determined from the config fil
Anybody who knows how to create such multiple instances.
Hmm, so there are two processes, but they want the processes to be able to communicate with each other via more than one pair of sockets? i.e. you could have two (or more) TCP socket connections between the two processes, and/or two (or more) pairs of UDP sockets sending packets back and forth.
If my above paragraph is correct (i.e. if I haven't misunderstood the request), that is certainly possible, although it's not terribly obvious what advantage you'd gain by doing it. Nevertheless, what you'd need to do is have each instance of your application create multiple sockets (either by socket()+bind() for a UDP socket, or by socket()+bind()+listen()+accept() for accepting an incoming TCP connection, or by socket()+connect() to initiate a TCP connection to the other program instance.
The tricky part with managing multiple sockets is handling the waiting correctly. With just one socket you can often get away with using the default blocking I/O semantics, and that way you can end up treating the socket something like a file, and just let each send() or recv() operation (etc) take however long it needs to take to complete before it returns to your calling function.
With more than one socket, on the other hand, you typically want to be able to respond to data on any of the sockets that are ready, which means that you can't just block waiting on any one particular socket, because if you do that, you may end up stuck waiting for a long time (potentially forever!) before that blocking call returns, and in the meantime you are unable to handle any data coming in from any of the other sockets. (The problem becomes particularly obvious when one of the connections is to a computer whose plug was just pulled, as it will typically take the TCP stack several minutes to figure out that the remote computer has gone away)
To deal with the problem, you'll typically want to either use non-blocking I/O and a socket-multiplexing call (e.g. poll() or select() or kqueue()), or spawn multiple threads and let each thread handle a single socket. Neither approach is particularly easy -- the socket-multiplexing approach works well once you get the hang of it, but the multiplexing calls' semantics are somewhat complex, and it takes a while to understand fully how it is intended to work. Non-blocking I/O complicates things further, since it means your code has to correctly deal with partial reads and writes. The multithreading approach seems simpler at first, but it has its own much larger and more subtle set of 'gotchas' (race conditions, deadlocks) that can cause much pain in the long run if you aren't very careful about what the threads are doing and how.
ps Since you're in a Unix environment, a third possible approach would be to fork() a child process for each socket. This would be similar to the multithreading approach, except a bit safer since your "threads" would actually be processes and each would have their own separate memory space, and thus they'd be less likely to trip over each other while doing their work. The downside would be higher memory usage, and also it becomes a bit harder (and slower) for the processes to communicate with each other due to the process space separation.