Subject: Re: [boost] [c++std-parallel-2020] Re: We need a coherent higher level parallelization story for C++ (was [thread] Is there a non-blocking future-destructor?)
From: Peter Dimov (pdimov_at_[hidden])
Date: 2015-10-14 12:47:51
Lawrence Crowl wrote:
> First, std::async and std::future have no flaws for what they were
> designed to do. They were designed to provide relatively easy parallelism
> for small core systems under the fork-join model.
I thought that it was well established by now that std::async isn't very
useful for fork-join (if at all.)
First, it's required to execute each (asynchronous) task as-if in a new
thread with respect to thread locals, which in practice means "in a new
thread".
Second, it has to allocate the future shared state.
These two non-flaws kill its performance for fork-join parallelism.
The first one is not related to its use of std::future, but the second one
is. Given an interface of the form
X x;
async_result<R> r;
std::async( r, [&]{ return x.f(); } );
with async_result being non-copyable and non-movable and ~async_result
blocking on the async task, the task can store its result directly into r,
without needing to allocate.