[rabbitmq-discuss] So who's Judy and what does she have to do with the High Speed TCP driver for Erlang?

[Matthias Radestock]

On 18/07/12 16:04, Tony Garnock-Jones wrote:
> It looks like you could summarise the improvements as: instead of
> requesting an exact number of bytes in the reader loop, and waiting for
> that exact number to be delivered, when you need more input you set
> active_once, which feeds you whatever's available. You take what's
> given, and if it's (more than) enough, consume it up until it is no
> longer sufficient, at which point you set active_once again.
>
> Even shorter: the new code avoids use of async_recv and instead sets
> active_once when it wants more input, until it has enough to work with.
>
> Is that right?

Sounds plausible, though I think there's more to it. Note the 
tune_buffer_size, for example.

> Can you say anything about how much better things are this way? E.g.
> some percentage improvement to throughput, presumably?

I don't remember and it would take me a while to dredge up the information.

Matthias.