[nanomsg] Re: why memcpy data to be sent?
- From: Alex Elsayed <eternaleye@xxxxxxxxx>
- To: nanomsg@xxxxxxxxxxxxx
- Date: Wed, 29 Oct 2014 14:39:02 -0700
Paul Colomiets wrote:
> Hi Alex,
>
> On Wed, Oct 29, 2014 at 12:52 PM, Alex Elsayed
> <eternaleye@xxxxxxxxx> wrote:
>> The points people have been bringing up against it - zero-copy not being
>> worth it under 512K, etc - are mostly from the
>> crossing-protection-domains type.
>>
>
> So do you suggest that zero-copy between _layers_ make sense on larger
> or smaller messages?
>
Ownership transfer between layers can have basically no runtime overhead,
since you can enforce it at the language level (C++ unique_pointer<T>, Rust
...basically everything except &references); thus it can be beneficial for
any message smaller than a pointer. With inlining, it can be _actually_
free, since you don't even need to marshal the pointer into function
arguments.
However, the downsides mostly manifest in languages like C, where there
isn't One True Language-Blessed Destructor Protocol. Because without that,
the fact that ownership transfer also transfers the responsibility to free
memory becomes a liability, and requires additional APIs like a custom free
callback.
Other related posts:
