Benchmark: NVML API vs sysfs read for GPU NUMA node ID retrieval

gpu_numa_bench.rs

#![cfg(feature = "gpu-topology")]

use nvml_wrapper::bitmasks::InitFlags;
use nvml_wrapper::Nvml;
use std::path::Path;
use std::time::Instant;

fn read_from_file<T: std::str::FromStr>(path: &Path) -> Option<T> {
    std::fs::read_to_string(path)
        .ok()
        .and_then(|s| s.trim().parse().ok())
}

fn main() {
    let nvml = Nvml::init_with_flags(InitFlags::NO_GPUS).expect("Failed to init NVML");
    let count = nvml.device_count().expect("Failed to get device count");
    println!("Found {} NVIDIA GPU(s)\n", count);

    const ITERATIONS: u32 = 100_000;

    for i in 0..count {
        let gpu = nvml.device_by_index(i).expect("Failed to get device");
        let pci_info = gpu.pci_info().expect("Failed to get PCI info");
        println!("GPU {}: bus_id={}", i, pci_info.bus_id);

        // Benchmark: NVML API call (new path)
        let start = Instant::now();
        for _ in 0..ITERATIONS {
            let _ = std::hint::black_box(gpu.numa_node_id());
        }
        let api_elapsed = start.elapsed();

        // Benchmark: sysfs read (old path)
        let bus_id = pci_info.bus_id.to_lowercase();
        let fixed_bus_id = bus_id.strip_prefix("0000").unwrap_or(&bus_id);
        let numa_path = format!("/sys/bus/pci/devices/{}/numa_node", fixed_bus_id);
        let start = Instant::now();
        for _ in 0..ITERATIONS {
            let _ = std::hint::black_box(read_from_file::<usize>(Path::new(&numa_path)));
        }
        let sysfs_elapsed = start.elapsed();

        println!("  NVML API (numa_node_id):  {:?} ({} iters)", api_elapsed, ITERATIONS);
        println!("  sysfs read:               {:?} ({} iters)", sysfs_elapsed, ITERATIONS);
        println!(
            "  Speedup: {:.1}x faster",
            sysfs_elapsed.as_nanos() as f64 / api_elapsed.as_nanos() as f64
        );
    }
}