feat(deadpool): cache num_cpus for better init performance

[PureWhiteWu]

On Linux environment, calling num_cpus::get_physical() costs about 1-2ms per time, and in our scenario, we need to create hundreds to thousands of Pools dynamically(our services normally have hundreds to thousands of instances).

Thus, it will cost a really long time to create the pools, which lead to service hang or request timeout.

This PR caches the result of num_cpus::get_physical to avoid calling it every time, which can greatly improve the new performance of Pool.

Also, in container environment, we should call num_cpus::get() instead of num_cpus::get_physical(), but since it may be a breaking change, I only added an TODO for it.

PTAL, thanks!

[bikeshedder]

Please also add an entry to the CHANGELOG.md under the ## [Unreleased] describing the change.

[PureWhiteWu]

CHANGELOG updated.

[PureWhiteWu]

Oh, LazyLock requires Rust 1.80, but the MSRV is 1.75 now.

Should I switch it to lazy_static or bump the MSRV?

[PureWhiteWu]

I've switched to lazy_static for MSRV compatibility.

[bikeshedder]

Sticking with the current MSRV of 1.75 is probably the better choice.

btw. regarding the TODO refering to num_cpus::get() instead of num_cpus::get_physical() I feel like this is a non-breaking change but rather fixes wrong behavior, no?

[bikeshedder]

When running a container with 4 CPUs on an Ampere system with 192 cores I would not want to end up with a default pool size of 768 (192*4), would I? 🤔

[bikeshedder]

Ah, I remember reading the documentation and went with get_physical as logical cores aren't very useful in a lot of workloads. I ended up using physical CPUs for that very reason. tbh. I didn't think of cgroups and containers back then. So yes, it should probably be changed to something like num_cpus::get() * 2 so the default stays the same for most of the users and does the right thing in containerized environments, too.

[bikeshedder]

I'm happy to merge this. I just would like to make a decision on that TODO now. It has the potential of staying in the source code for the next years to come.

My vote would be to go with num_cpus::get() * 2 and call it the day.

[bikeshedder]

@PureWhiteWu I just merged this and created a separate PR for the default max pool size:

• Fix default pool max size #441