Subject: Re: [boost] [afio] AFIO review postponed till Monday
From: Ben Pope (benpope81_at_[hidden])
Date: 2015-07-22 17:37:10
On Thursday, July 23, 2015 12:01 AM, Niall Douglas wrote:
> On 22 Jul 2015 at 11:12, TONGARI J wrote:
>
>>> Quite a few people have disliked (a) the choice of future
>>> continuations over ASIO's async_result pattern and (b) the batch API.
>>> Those two observations have come up repeatedly - Bjorn and Robert on
>>> this list have both publicly found issue there, and neither was alone
>>> in their opinion.
>>>
>>> I'm not dropping futures in favour of async_result - I don't think
>>> that helps ease of use because in file i/o you really do want strong
>>> i/o ordering, and you usually don't care about ordering much for
>>> network i/o. Forthcoming C++ 1z coroutines are also futures based,
>>> and that decision is not going to be reversed now. Futures are our
>>> future as it were.
>>>
>>
>> async_result is more flexible and it allows the use of future, e.g.
>> `s.async_xxx(..., asio::use_future)`, and you can customize your
>> afio::use_future to use your lightweight futures.
>> What's your real concern of async_result? The verbosity of specifying
>> afio::use_future?
>> Since you mentioned C++ 1z coroutines, the asio way actually allows
>> zero-overhead adaption without using a future, e.g. see:
>> https://github.com/jamboree/act
>
> A good summary of why those who prefer async_result do so, thank you.
>
>> Is any performance reason in preferring strict future API to the callback
>> style?
>
> Historically futures were slower, and Chris said so repeatedly in
> N-papers to WG21.
>
> I believe lightweight futures have eliminated the performance gap. I
> also have an extension to the Concurrency TS in that you can add as
> many const lvalue ref continuations as you like to lightweight
> futures, not just a single rvalue ref continuation as per the TS.
>
> This effectively implements a single shot signals and slots
> implementation - you can install as many completion handlers as you
> like per future.
>
> None of this answers your question of course. I will say that file
> i/o is often 10x-1,000,000x slower than network i/o, and the
> performance of handlers vs continuations is correspondingly
> unimportant.
>
> Futures gives you the option of monadic programming, and they enforce
> strict operation ordering very succintly. They are the right choice
> for file i/o, just as async_result is the right choice for network
> i/o.
>
> Though, to be honest, I suspect lightweight futures will make
> async_result not as useful as previously.
To clarify; I think you're saying that your lightweight futures are
close enough to free that you would prefer to implement your library
with them as opposed to callback hell; and so, presenting the nicer
future API poses no significant further performance penalty whilst
presenting a nicer syntactic model.
Personally I'm really interested in your lightweight futures for the
same reason. Callback hell sucks; without profile information to the
contrary, I'd rather just write code the nice way.
Ben