Re: [RFC] random: use per lcore state

[Stephen Hemminger <stephen@networkplumber.org>]

On Wed, 6 Sep 2023 19:54:32 +0200
Morten Brørup <mb@smartsharesystems.com> wrote:

> > 
> > -	idx = rte_lcore_id();
> > +	seed = __atomic_load_n(&rte_rand_seed, __ATOMIC_RELAXED);
> > +	if (unlikely(seed != rand_state->seed)) {  
> 
> Please note that rte_rand_seed lives in a completely different cache
> line than RTE_PER_LCORE(rte_rand_state), so the comparison with
> rte_rand_seed requires reading one more cache line than the original
> implementation, which only uses the cache line holding
> rand_states[idx].
> 
> This is in the hot path.
> 
> If we could register a per-thread INIT function, the lazy
> initialization could be avoided, and only one cache line accessed.

Since rte_rand_seed rarely changes, it will get cached by each cpu.
The problem was before there was prefetch cache overlap causing false
sharing.