On 09/14/2010 07:25 PM, Steven McCoy wrote: > The latter is one sequence number in size otherwise you can get the > situation that receivers are asking for re-transmission of packets that > haven't been sent yet. Timer based TX flushing puts transmission time > into the receiver thread for bus transports which will only be adverse > for performance.
Ok. I see. But we still have kernel tx buffer, so when kernel signals that buffer if full, we can sleep for a while till the data are actually pushed to the wire. With 128kB transmit buffer and 1Gb/s transmit rate it takes ~1ms to flush the data. What if I make the built-in 0MQ timers more precise? To the level of milliseconds? That could solve the problem, no? We can then compute the value for SO_SNDBUF so that it takes 1ms to flush. (The thing I am trying to avoid is implementing precise timers on 0MQ level. It's a mess I would rather not maintain.) Martin _______________________________________________ zeromq-dev mailing list [email protected] http://lists.zeromq.org/mailman/listinfo/zeromq-dev
