I'm working on a C++ socket server using I/O completion ports. I create one I/O port and two worker threads per processor as some articles recommend in order to handle the overlapped I/O activity
From what I've read only one of these worker threads should be waking up and handling an I/O request at any time. So what I would expect when one client connects (only) and sends something is only one thread to wake up and handle that receive, but when I try to debug it I can see that multiple threads wake up and try to handle the same operation.
Is my assumption wrong?
Related
I'm writing a low-latency program in C++ which receives data from a source, processes the data and sends to a target via a TCP socket. I have a separate thread for all these 3 modules, receiver thread, processor thread, sender thread. All these threads are communicating with lock-free queues.
Do you think that sending the message directly and not using the queue for the sender part would give lower latency? Does it affect performance stability?
Thanks
If the three threads are pinned to different physical cores, having a separate sender thread would give lower latency than processor thread doing the send operation, specially if there are retries happening in the send process. Even if it is a best effort async send, you could still save the marginal time it takes to write to the socket.
I must develop a simple C++ command line client/server chat application. This application must provide a basic multiple two-partecipants chat-room implementation. Is it possible to combine IO/Multiplexing (select() syscall) with POSIX threads?
I mean I want to create a TCP server which handles multiple clients with select() and when a client wants to chat with another one the servewr creates a separate thread , that uses IO/Multiplexing (select() syscall) , to handle the communication between the two clients.
Is this a good idea? How could I do otherwise?
A crude attempt at an architecture...
Structure your application as two sets of threads (a set might be composed of just one thread).
One set minds the TCP connections, each TCP connection is assigned to one of the threads in the set, the thread just runs forever polling the connections assigned to it (incoming messages) and polling a (per-thread) from-logic queue (outgoing messages)
The other set minds the logic/session. Each session is assigned to a specific thread. Each thread just runs forever polling the (per-thread) from-network queue (incoming messages).
The network thread-set, receives messages and post them to the right logic queue [assumes there's a way of mapping connections to internal logic sessions]. It polls its from-logic queue to get the outgoing messages and send them.
The number of network threads is bound, and it does not depend on the number of connections.
The logic thread-set, receives requests from the network in its queue and handles them within a given session state and (perhaps) post back messages to the be sent out (sent out by the network threads)
The number of logic threads is bound, and it does not depend on the number of sessions.
Is there a OS portable way of interrupting a blocking accept? I have a multi-threaded deamon/service that needs to gracefully close all listening network hosts when the deamon/service gets a shutdown signal. I see that some say you should use non-blocking sockets or select with a time-out - but won't these degrade performance as my application should be as fast as possible? The deamon runs in the main thread while each of the listening network hosts run in their own thread. The problem now is that accept wait indefinitely if there is no network traffic on the listening network host's socket. If I should use signals, then is there an example somewhere of using signals to interrupt accept?
The solution here is to not call accept when there's nothing to do. Just use non-blocking select or poll to wait until there's something to accept, then accept at that point. Unless you create a really tiny timeout there won't be any performance implications of waking up from the non-blocking call and going back to wait on that socket again.
I am faced with a design issue regarding thread synchronization in C++, Windows.
I am writing a server application that starts one listening thread, which should stay active the whole time while the server is up.
When the listening thread gets a connect request, it opens a CONTROL socket and starts a new control thread.
This thread is used to send control data between server and a client, initializing server and all the background software to specific client data and starting data processing.
If the initialization (via control socket) is successful, the control thread will open a new socket, DATA socket, which is then used to pass data from server to client. It will also start two new threads, one which is sending on this new, DATA socket, and the other, which is receiving on the CONTROL socket, waiting if the client wants to terminate connection.
When client terminates connection ungracefully, by terminating application without the call to function which sends the server message to close the connection, here is what should happen:
Any of the threads in execution can detect this event. They will get some sort of error (WSAECONNRESET) while sending or receiving on DATA/CONTROL socket and should then signal all the other threads that they should stop executing (except for the server listening thread).
Which is the most natural way to achieve this type of behavior?
(I am using winsock (winsock2.h) for networking, and standard windows api (windows.h) for threading)
If you're writing a multi-threaded winsock server, you should be looking into IO completion ports. Using an IO completion port is the most scalable way to write a network service on the windows platform.
IO completion port based winsock servers use asynchronous communication, so instead of blocking on a socket, your threadpool receives completion packets when something interesting happens.
In any case, you'll be using WSARecv. When WSARecv returns non zero, call WSAGetLastError(). If you don't have WSA_IO_PENDING, then switch on the error and look for the winsock error code you're interested in.
The winsock error code WSA_OPERATION_ABORTED indicates that a socket has closed, although there are others (e.g. WSAECONNABORTED).
Would suggest a good text on the subject (e.g. Windows via C/C++).
You can use WSAEventSelect() function to associate event object with socket and create one event object for your events, then use these event objects in WaitForMultipleObjects() function, so your thread can wait for socket events and your custom events.
I need a client networking thread to be able to respond both to new messages to be transmitted, and the receipt of new data on the network. I wish to avoid this thread performing a polling loop, but rather to process only as needed.
The scenario is as follows:
A client application needs to communicate to a server via a protocol that is largely, but not entirely, synchronous. Typically, the client sends a message to the server and blocks until a response is received.
The server may process client requests asynchronously, in which case the response to client
is not a result, but a notification that processing has begun. A result message is sent to to the client at some point in the future, when the server has finish processing the client request.
The asynchronous result notifications can arrive at the client at any time. These notifications need processed when they are received i.e. it is not possible to process a backlog only when the client transmits again.
The clients networking thread receives and processes notifications from the server, and to transmit outgoing messages from the client.
To achieve this, I need to to make a thread wake to perform processing either when network data is received OR when a message to transmit is enqueued into an input queue.
How can a thread wake to perform processing of an enqueued work item OR data from a socket?
I am interested primarily in using the plain Win32 APIs.
A minimal example or relevant tutorial would be very welcome!
An alternative to I/O Completion Ports for sockets is using WSAEventSelect to associate an event with the socket. Then as others have said, you just need to use another event (or some sort of waitable handle) to signal when an item has been added to your input queue, and use WaitForMultipleObjects to wait for either kind of event.
You can set up an I/O Completion Port for the handles and have your thread wait on the completion port:
http://technet.microsoft.com/en-us/sysinternals/bb963891.aspx
Actually, you can have multiple threads wait on the port (one thread per processor usually works well).
Following on from Michael's suggestion, I have some free code that provides a framework for IO Completion Port style socket stuff; and it includes an IOCP based work queue too. You should be able to grab some stuff from it to solve your problem from here.
Well, if both objects have standard Windows handles, you can have your client call WaitForMultipleObjects to wait on them.
You might want to investiate splitting the servicing of the network port off onto its own thread. That might simplify things greatly. However, it won't help if you just end up having to synchonize something else between that new thread and your main one.