[PATCH grout] api: optimize send small payloads

[Robin Jarry]

Morten Brørup, Jul 06, 2026 at 12:41:
>> Did you measure any improvement after this patch?
>
> I did not profile it, no. The send() syscall is quite expensive, so
> combining two calls into one should give Grout API transactions
> a measurable performance boost.

OK, in that case (unless we can show a reasonable improvement), I would
prefer to keep the code simpler.

You may save one write() call, but the other side needs to read() twice
anyway, once to get the header (which tells the payload length) and
a second time for the payload (if any).

> The changes related to the bufferevent_write() lib call were added
> mainly for symmetry. So I set a much lower merge threshold. I think
> the performance difference negligible. BTW, my changes here also fixes
> a bug: The payload was unconditionally written, regardless of writing
> the header failed.

The main concern would be on the daemon side but as you pointed out,
this is already dealt with in the bufferevent_write() abstraction. Let's
not add another level of (conditional) buffering.

We could fix that extraneous write, but I don't think it causes any
issue. If one write fails, we'll get notified in event_cb() and grout
will close the connection anyway.

> Honestly, I just started familiarizing myself with Grout a few weeks
> ago, and stumbled across this (inefficiency of calling send() twice)
> while trying to learn about Grout's API mechanisms. So I thought I'd
> add this simple optimization.



-- 
Robin

> Batteries not included.