Kernel threading does scale to the number of threads applications need these days.

What can be slow are context switches, but they aren't slow in absolute terms. The vast majority of applications, including Web servers, are perfectly fine with 1:1 threading.

It doesn't really scale. Kernel level threads are quite expensive in the terms of memory. Erlang's processes take up few hundred bytes. It's rare to have dozen thousands of kernel-level threads running on Linux, while it's quite common for Erlang servers to have that many processes.

Anyone know if this is better in Alpine Linux with musl libc and smaller stacks?

It's not really up to a distribution to lessen this cost. Large part of it is what task descriptor in the kernel takes.

Also, if nothing else, you'll run out of PID numbers, as they're usually still 16 bits, even today, though there was a kernel compilation option to change that, from what I remember.

It can be changed at runtime. From proc(5), system wide limits:

/proc/sys/kernel/pid_max

> PID_MAX_LIMIT, approximately 4 million

/proc/sys/kernel/threads-max

> FUTEX_TID_MASK (0x3fffffff), [approximately 1 billion]

For per-process limit, increase RLIMIT_NPROC.

> Large part of it is what task descriptor in the kernel takes.

4K for kernel stacks, only 2K with future work. That's really not much space at all if you're doing anything interesting with those threads.

> Also, if nothing else, you'll run out of PID numbers, as they're usually still 16 bits, even today

Not for a very long time.

I wrote a server application that runs with about a million goroutines and performs quite well. It's not a webserver, but it responds to "requests" within hundreds of micros. Surely this would not be possible with OS threads.

And the go implementation is also significantly broken, as evidenced by the namespace issue a while ago. https://news.ycombinator.com/item?id=14470231

Then look at what Clojure is capable of.

Threads do scale. On Linux, O(1) scheduler solved this non-issue a long time ago.

> Threads do scale. On Linux, O(1) scheduler solved this non-issue a long time ago.

yup they do.

till you start making sure that your code doesn't end up with deadlock, data-corruption, races, performance issues due to lock-contention etc. etc. designing efficient locking schemes is notoriously hard alternating between:

- too coarse grained : resulting in serializing activities which could have (should have) proceeded in parallel, thereby sacrificing performance and scalability.

or

- too fine grained: with space+time for lock operations sapping performance, error recovery and not to mention understanding etc. etc.

In the former we have the dragons of deadlock and livelock roaming freely, and in the latter we have race conditions. Somewhere in between is a razor's edge which is both efficient and correct.

Almost always, things start with ‘one big lock around everything’ and the vague hope that performance might not be abysmal. When that is dashed, big lock gets broken up, and the prayer is repeated. Each iteration increasing complexity and decreasing lock-contention, and hopefully with some luck, modest performance gain as well.

remember this:

What do we want ?

Now !

When do we want it ?

Fewer race conditions !

have fun :)

I see your design is lacking and your fud quotient is high. Good on you!

Yep. This was my reaction as well. Threads are fine.

Having high numbers threads and switching between threads are different things. There is still a huge constant in front of that O(1) scheduler that makes it unattractive.

Spin up 100k threads on linux vs 100k actors in erlang.