UDP套接字缓冲区溢出检测

Question

I'm developing cross-platform tool that captures multiple udp streams with various bit-rate. boost::asio is used for networking. Is there any way to detect the situation when the udp buffer was full and data loss on socket could take place? The only way I can see now is reading /proc/%pid%/net/udp, but it's not aplicable for windows as you know :). Also I'd like to use boost features for it if possible.

Answer 1

If you need this capability, you have to code it into the protocol you are using. UDP is incapable of doing this by itself. For example, you could put a sequence number in each datagram. Missing datagrams would correspond to missing sequence numbers.

Answer 2

I've just hit the same issue (although for me Linux-specific), and despite the question being old might as well document my findings for others.

As far as I know, there are no portable way to do this, and nothing directly supported by Boost.

That said, there are some platform-specific ways of doing it. In Linux, it can be done by setting the SO_RXQ_OVFL socket-option, and then getting the replies using recvmsg(). It's poorly documented though, but you may be helped by http://lists.openwall.net/netdev/2009/10/09/75 .

One way to avoid it in the first place is to increase the receive-buffers (I assume you've investigated it already, but including it for completeness). The SO_RCVBUF options seems fairly well-supported cross-platform. http://pubs.opengroup.org/onlinepubs/7908799/xns/setsockopt.html http://msdn.microsoft.com/en-us/library/windows/hardware/ff570832(v=vs.85).aspx OS:es puts an upper limit on this though, which an administrator might have to increase. On Linux, IE it can be increased using /proc/sys/net/core/rmem_max.

Finally, one way for your application to assess it's "load", which with large input-buffers might serve for early detection of overloading, could be to introduce a timestamp before and after the async operations. In pseudo_code (not boost::async-adapted):

work_time = 0
idle_time = 0

b = clock.now()
while running:
  a = clock.now()
  work_time += a-b
  data = wait_for_input()
  b = clock.now()
  idle_time += b-a
  process(data)

Then every second or so, you can check and reset work_time / (work_time+idle_time) . If it approaches 1, you know you're heading for trouble and can send out an alert or take other actions.