[zeromq-dev] 0MQ for hard real time applications: the inproc communication

[Martin Hurton]

> My main question at the moment is about the current implementation of INPROC
> transport.
> Digging it seems like the actual implementation is the class pipe_t, am I
> right?

Yes, producer side writes the message to the pipe and consumer reads
it from it. The pipe is implemented as lockfree on some platforms.
There is some overhead when the consumer is sleeping, as a "wakeup"
control message has to be delivered.

> I have done some benchmark using the perf tool (inproc_lat) and I get an
> average latency of 8 usec.
> It is not bad, but using lockfree FIFO I can achieve times that are TWO
> orders of magnitude lower.

You mean 80 ns?

> I know that the beauty of ZeroMQ it is that it doesn much more for me, but I
> would like to understand better the current implementation of INPROC sockets
> to see if I can contribute with a more efficient implementation.

The lock-free fifo is implemented in ypipe.hpp.
This is used to implement pipes (message queues). These are
bi-directional and implement internal signalling which is used for
flow control (so that producer is woken up when more messages can be
pushed and consumer is woken up when another message can be pulled).
The inproc is implemented using a pipe.
Socket.cpp contains the code contains plumbing + implementation of API layer.

I hope this can help you start with your changes.
Do not hesitate to ask should you have more questions.

- Martin