[zeromq-dev] idea for perf enhancement in user code

[Martin Sustrik]

On 12/14/2011 11:33 AM, Chuck Remes wrote:
> I have a protocol (like majordomo) that uses message parts to send
> some of the same "boilerplate" data on every message. Currently I
> allocate a new zmq_msg_t for each of these "frames" and let the
> library handle the deallocation, but it occurred to me that this
> might not be a good idea.
>
> What if I allocate the zmq_msg_t once and, before passing it to
> zmq_send/zmq_sendmsg, I call zmq_msg_copy on it to increase its
> refcount. The library returns another zmq_msg_t to me in the buffer I
> gave it but it let's me avoid the memcpy call for the "data" (which
> may be a negligible cost for small data but could be significant for
> large buffers).

Yes. That would work. You can have a boilerplate message part stored and 
send it this way:

zmq_msg_copy (tmp, boilerplate);
zmq_send (tmp);

> Now that I have typed this out, it would be nice if I could operate
> *directly* on the original zmq_msg_t and avoid the effort of even
> creating a destination zmq_msg_to for zmq_msg_copy to replace. What
> about a zmq_msg_increment_refcount(zmq_msg_t *src) api call to let me
> do this work directly?

In theory we can add something like zmq_msg_add_ref() to get that 
functionality.

The question is whether more explicit reference management using 
separate zmq_msg_t per reference doesn't lead to better coding style.

Martin