Re:Scipio: A Thread-per-Core Crate for Rust and Linux
by Joker_vD on Tuesday November 03, 2020 @08:24AM
(#24979389)
There is an easy solution to the colour problem: make all functions async. All of them. Of course, having code like this
is not fun, especially when you have to await the result of the arithmetic operators, and write all those now-redundant async and await, so let's also drop "async" and make "await" implicit--although we'll need something for non-awaiting. Let's call it "nowait". Now we can write code like this:
A-a-and we're back to multithreading, basically. So yeah, looks like cheap context switches is the solution.
[ Parent
| on HN ]
Re:Scipio: A Thread-per-Core Crate for Rust and Linux
by gpderetta on Tuesday November 03, 2020 @11:00AM
(#24980780)
Agree :) . Semantically there is no difference, but in practice async functions are compiled differently than normal functions. what you are proposing in practice is to do a global CPS transform and possibly reconstruct back the original function when the continuation is not needed (i.e. it is always invoked in strict LIFO order). There are functional languages where this is the norm, but I don't think this is appropriate for a system language like rust (it would greatly complicate interop with C for example).
My idea is to make the continuation explicit and CPS transform all and only the functions that have a continuation as parameter (and any generic function with a type that is a continuation or contains a continuation). Fall back to dynamic stacks if the continuation escapes.
It is always continuations all the way down.
Re:Scipio: A Thread-per-Core Crate for Rust and Linux
by Joker_vD on Tuesday November 03, 2020 @11:40AM
(#24981233)
No, it's not necessary to perform a CPS transformation. After all, the OS manages to pre-emptively schedule any non-cooperating processes, right?
So, leaving aside the problem of interacting with the OS for I/O, you can do interleaved CPU-bound in a single thread, by instrumenting your code with basically an instruction counter. When it grows a bit too large, stop, reset it, and switch to another task. Erlang does this, although being a functional language, it counts only function calls (no other way to make a loop than to (tail-)call itself). No CPS needed.
Re:Scipio: A Thread-per-Core Crate for Rust and Linux
by gpderetta on Tuesday November 03, 2020 @12:03PM
(#24981468)
Yes of course, a stack per thread is always an option.
Re:Scipio: A Thread-per-Core Crate for Rust and Linux
by glaubercosta on Tuesday November 03, 2020 @08:44AM
(#24979564)
Maybe I am part of the problem by emphasizing context switching too much, but as penberg noticed, there is more to thread per core than context switching.