[ruby-core:88484] Re: [Ruby trunk Feature#13618] [PATCH] auto fiber schedule for rb_wait_for_single_fd and rb_waitpid

[Eric Wong <normalperson@yhbt.net>]

Koichi Sasada <ko1@atdot.net> wrote:
> On 2018/08/14 9:42, Eric Wong wrote:
> > I am thinking of adding preemption support to this feature for
> > compatibility with 1.8
> 
> So that "auto-fiber" proposal is to provide green threads like Ruby 1.8?

Yes; this was always the idea.

> Like:
> ```
>   model 1: Userlevel Thread
>     Same as traditional ruby thread.
> ```
> in thread.c comment (I wrote 13 years ago!).
> 
> I don't against this idea, but I think it is hard to select these options by
> Ruby programmers. I think changing Thread implementation model from native
> thread (1:1 model) to green thread mode (1:N model) is better for Ruby
> programmers.

No, I want to keep current Thread 1:1 model because it is useful
for filesystem operations and some CPU/memory intensive tasks (zlib).

Changing "Thread" now to 1:N would also break code written for 1.9..2.5

1:N model is good for most network operations (high latency, low throughput).

> To change them, we need to discuss pros. and cons. of them carefully. There
> are several good points (the biggest advantage of 1:1 model is friendly for
> outer libraries) but are bad points (1:1 model has performance penalties,
> and recent glibc malloc arena issues and so on).

Agreed.

> I don't think it is a good idea to choose such internal implementation by
> Ruby programmers. ... easy?

I think it is necessary to give that control to programmers.
We can educate them on pros and cons of each and to use them
in combination.

```

1:1   + C-ext parallelism for SMP systems (example: zlib)
      + filesystem parallelism (no non-blocking I/O on FS ops)
      + external library compatibility
      + compatibility with 1.9..2.5 code

      - high memory use (malloc arenas, kernel structs)
      - kernel resource limitations (Process::RLIMIT_NPROC)


1:N   + C10K problem (or C100K :P)
      + compatibility with old 1.8 designs
      + low memory and resource use
        (malloc and kernel never sees extra data structures)

      - cannot take advantage of SMP or multiple filesystems alone
```

The key is a programmer may combine 1:1 and 1:N for different
parts of the program flow.  So where the program is bottlenecked
on filesystem, it can rely on 1:1 behavior, but when the program
is waiting on slow network traffic, it can rely on 1:N behavior

For example: I have 4 filesystems, I might have 32 native threads
(8 threads/FS to keep kernel IO scheduler busy).
But I will still serve 100K network clients with 100K 1:N threads
and that can even use 1 native thread.