I'm writing an application using Mina which continuously creates a large number of short lived UDP sockets with different local ports but the same destination address/port.
What I need to be able to do is create a UDP socket with either a unique or ephemeral local port, send one packet to the same destination on the same port, wait for a response, then close the socket. Do this, say, 10 million times. Using Mina 2.0, what I've done is, using the same NioDatagramConnector, call connect() again and again. I would use NioDatagramAcceptor.newSession() except that it's not the remote address that varies here, it's the local address. What I've found is that connect() eats up about 36% of the total application CPU time, and that the process itself pins the CPU to 100% ... which would be fine, except I'm only seeing about 3000 round trips / second on trivial 20 byte messages. I know that's a lot of system calls, but even so, if I were writing this in C/C++, I think I could do much better than 3000 round trips/second. At least 10K, 20K round trips/second should be readily achievable. Of course I need to rate limit so that only < 64K fds are being used at a time. I'm guessing that one of the performance implications with connect() is that it spins up machinery to perform an IoFuture callback ... but with UDP sockets, it doesn't seem necessary to do async notification on connect() as no packets are actually being exchanged -- just reservation in the kernel for the datagram socket. Using the java nio DatagramChannel directly, I could call connect() myself, and I'm sure it would be very fast. But I see no (obvious) way to construct an IoSession from an already existing and bound DatagramChannel. Advice? -- View this message in context: http://www.nabble.com/performance-of-NioDatagramConnector.connect%28%29-tp14683147s16868p14683147.html Sent from the Apache MINA Support Forum mailing list archive at Nabble.com.
