Very simple performance experiment

Results

Warm-up

No obvious warm-up effect across the reps, but:

• No pre-iteration: 1.3ms; 1.4ms
• Pre-iterate: 1.2ms; 1.3ms

Using

    if items.iter().any(|i| i.id[0] >= u64::MAX) {
        return ([0, 0, 0, 0], Duration::MAX);
    }

Build

2sd mean; p99 (m3)

• debug: 4.3ms; 4.6ms
• release: 1.3ms; 1.4ms
• release + abort on panic: 1.3ms; 1.4ms
• release + optimised profile + RUSTFLAGS="-C target-cpu=native": 1.3ms; 1.4ms

Timing the right thing

2sd mean; p99 (m3)

• Including init data and destruction: 83ms; 93ms
• Destruction only: 1.4ms; 1.8ms
• Neither: 1.4ms; 1.6ms
• Single allocation: 1.3ms; 1.4ms

Loop style

2sd mean; p99 (m3)

• for_each: 1.3ms; 1.4ms
• for i in items: 1.3ms; 1.4ms
• for i in 0..items.len(), for i in 0..SIZE, etc: : 1.3ms; 1.4ms
• and with black_box: 2.7ms; 3.5ms

Hard to force the bounds checks!

Overflow checks

Manual overflow checks didn't make much difference

Parallelism

Using atomic u64s for result shows the same times.

• Single thread: 1.3ms; 1.4ms
• 10 threads, bad atomics: 19.9ms; 26.3ms
• 10 threads, good atomics: 0.9ms; 0.9ms
• 2 threads, good atomics: 0.9ms; 1.0ms
• 10 threads, interleaved: 1.3ms; 1.7ms
• Rayon (par_iter): 43.1ms; 53.6ms

Data-oriented approach

• OO: 1.3ms; 1.4ms
• split ids from values: 1.1ms; 1.2ms
• split ids: 0.8ms; 0.9ms

Iteration style

Based on data-oriented, using indexing vs iterators is equivalent, seems to be some small but consistent benefit to indexing (0.81ms vs 0.83ms). With black_box on indices, mean is 0.97ms.