You can minimize sharing data between threads because it's easier to have data affinity with threads (ie only thread A will read or write to a piece of data). You can still access that data from multiple modules because the whole thread is never blocked waiting for IO (because of async). An extreme example is nodejs, where you only have one thread, can concurrently do thousands of things and never have to coordinate (ie via mutexes) data access.
It's not either or, you can combine the two. I've worked on a system that did real time audio mixing for 10000s of concurrent connections, utilizing >50 cores, mostly with one thread each. Each thread had thread-local data, was receiving/sending audio packets to hundreds/thousands of different IP addresses just fine without worrying about mutexes at all. Try that with tens of thousands of actual OS threads and the associated scheduling overhead.
Having data affinity to cores is also great for cache hit rates.
Here is part of the C++ runtime this is based on: https://github.com/goto-opensource/asyncly. I was the principal author of it when it was created (before it was open sourced).
> Each thread had thread-local data, was receiving/sending audio packets to hundreds/thousands of different IP addresses just fine without worrying about mutexes at all.
it doesn't sound they really sharing data with each other, it looks like your logic is well lineralizable and data localized, and you can't implement access to some global hashmap in that way for example.
> Try that with tens of thousands of actual OS threads and the associated scheduling overhead.
I run this(10k threads blocked by DB access) in prod and it works fine for my needs. There are lots of statements in internet about overhead, but not much benchmarks how large this overhead is.
> Here is part of the C++ runtime this is based on
yeah, I need one runtime on top of another runtime, with unknown quality, support, longevity and number of gotchas.
> it doesn't sound they really sharing data with each other, it looks like your logic is well lineralizable and data localized, and you can't implement access to some global hashmap in that way for example.
Yes, because data can have thread affinity. Data doesn't need to be shared by _all _ connections, just by a few hundred/thousand. This enables connections to be scheduled to run on the same thread so that they can share data without synchronization.
> I run this(10k threads blocked by DB access) in prod and it works fine for my needs. There are lots of statements in internet about overhead, but not much benchmarks how large this overhead is.
> yeah, I need one runtime on top of another runtime, with unknown quality, support, longevity and number of gotchas.
It's a library. It solved our problems at the time, years ago. It's still used in production and piping billions of audio minutes per month through it. You don't have to use it, I merely referred to it as an example. A similar library is proposed to be included in C++23: https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2023/p23...