NUMA Binding Causes Performance Degradation in Rayon Work-Stealing Demo

[fkysly]

Environment

• 32 cores (all physical cores + no hyperthreading)
• 2 NUMA nodes

Issue

When binding the following Rayon workload to specific NUMA nodes, performance significantly degrades.

fn execute(pool: &ThreadPool) {
    let given: NumIter = 
        (0..NUM_ITERATIONS)
            .into_par_iter()
            .map(|_| {
                let mut result = 0;
                for i in (2..100).rev() {
                    let mut is_prime = true;
                    for j in 2..i {
                        if i % j == 0 {
                            is_prime = false;
                            break;
                        }
                    }
                    if is_prime {
                        result = (result + i as usize) % 1000;
                    }
                }
                result
            })
            .sum();
}