pingpong/
pingpong.rs

1//! OSU-style micro-benchmark: point-to-point latency and bandwidth between
2//! ranks 0 and 1, plus an all-reduce timing across all ranks (to show the
3//! `O(log n)` collectives). Run with e.g. `mpiexec -n 2 ./pingpong`.
4
5use mpi::collective::SystemOperation;
6use mpi::traits::*;
7
8fn main() {
9    let universe = mpi::initialize().unwrap();
10    let world = universe.world();
11    let rank = world.rank();
12    let size = world.size();
13
14    if size >= 2 {
15        let latency_iters = 2000;
16        let bw_iters = 100;
17        let bw_bytes = 1 << 20; // 1 MiB
18
19        if rank == 0 {
20            // ---- latency: 8-byte round trip ----
21            let small = [0u8; 8];
22            // warm up
23            for _ in 0..100 {
24                world.process_at_rank(1).send(&small[..]);
25                let _ = world.process_at_rank(1).receive_vec::<u8>();
26            }
27            let t0 = mpi::time();
28            for _ in 0..latency_iters {
29                world.process_at_rank(1).send(&small[..]);
30                let _ = world.process_at_rank(1).receive_vec::<u8>();
31            }
32            let dt = mpi::time() - t0;
33            let lat_us = dt / latency_iters as f64 / 2.0 * 1e6;
34            println!("latency:   {lat_us:.2} µs (8 B, half round-trip, {latency_iters} iters)");
35
36            // ---- bandwidth: 1 MiB round trip ----
37            let big = vec![0u8; bw_bytes];
38            let t0 = mpi::time();
39            for _ in 0..bw_iters {
40                world.process_at_rank(1).send(&big[..]);
41                let _ = world.process_at_rank(1).receive_vec::<u8>();
42            }
43            let dt = mpi::time() - t0;
44            let mb = bw_bytes as f64 * bw_iters as f64 * 2.0 / 1e6;
45            println!("bandwidth: {:.1} MB/s (1 MiB, {bw_iters} iters)", mb / dt);
46        } else if rank == 1 {
47            let small = [0u8; 8];
48            for _ in 0..latency_iters + 100 {
49                let _ = world.process_at_rank(0).receive_vec::<u8>();
50                world.process_at_rank(0).send(&small[..]);
51            }
52            for _ in 0..bw_iters {
53                let (v, _) = world.process_at_rank(0).receive_vec::<u8>();
54                world.process_at_rank(0).send(&v[..]);
55            }
56        }
57    }
58
59    // ---- collective timing: all-reduce over all ranks ----
60    world.barrier();
61    let coll_iters = 1000;
62    let x = rank;
63    let mut acc = 0i32;
64    let t0 = mpi::time();
65    for _ in 0..coll_iters {
66        world.all_reduce_into(&x, &mut acc, SystemOperation::sum());
67    }
68    let dt = mpi::time() - t0;
69    if rank == 0 {
70        println!(
71            "all_reduce: {:.2} µs/op over {size} ranks ({coll_iters} iters), result {acc}",
72            dt / coll_iters as f64 * 1e6
73        );
74    }
75    world.barrier();
76}
pingpong/pingpong.rs

pingpong/
pingpong.rs
