search

Amanieu / atomic-rs

Generic Atomic<T> type for Rust
Apache License 2.0
218 stars 23 forks source link

False sharing will reduce the fallback locking concurency. #1

Closed dpc closed 6 years ago

dpc commented 8 years ago

I was just looking through this library (thanks for your work!) and I'm not sure if you are aware that most probably on most architectures the fallback array static SPINLOCKS: [AtomicBool; 1024] = array![ATOMIC_BOOL_INIT; 1024];, might be less effective than 1024-ways.

Typically the locking granularity is the the cacheline which is 64B on x86. As bools seems to be used in SPINLOCKS, each 64 consecutive entries in SPINLOCKS will occupy same cacheline so they will not be accessible on different CPUs at different times. So effectively this 1024-entry array works like a 16-entry one.

Sorry to bother you if you were already aware of that (16 seems like decent concurency level anyway). If not you can read more eg here:

https://www.ibm.com/developerworks/aix/library/au-aix-multicore-multiprocessor/ (ctrf+f "array of spinlocks") and generally google "spinlock false sharing".

Amanieu commented 8 years ago

I am aware of the lock contention issues, it's just that I haven't had time to work on Rust stuff lately. Still, thanks for opening an issue, I'll be sure to look into it at some point.

gnzlbg commented 6 years ago

Using a #[repr(align(64))] pub struct CacheLineAtomicBool(AtomicBool); [CacheLineAtomicBool; 1024] might be enough to fix this, but I don't recall if 64-byte alignment is supported.

Amanieu commented 6 years ago

Fixed in 0.4.0

gnzlbg commented 6 years ago

Nice, thanks!