lmflow 0.3.3

Dataflow-graph compute engine: a pure-Rust scheduler behind a stable C ABI
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159

//! Pure-Rust end-to-end throughput benchmarks.
//!
//! These benchmarks include graph input, scheduling, a private Rust pass-through
//! kernel, and Poller output. They complement `dispatch.rs`, which deliberately
//! removes Poller overhead to isolate scheduler dispatch cost.
//!
//! Run from `lmflow/core`:
//!
//! ```sh
//! cargo bench --bench throughput
//! ```

mod common;

use std::cell::Cell;
use std::hint::black_box;
use std::time::Duration;

use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion, Throughput};
use lmflow::{BufferData, Builtin, Graph, Packet, Timestamp};

const BATCH: u64 = 256;
const DT_U8: i32 = 0;

fn chain_yaml(depth: usize, threads: usize) -> String {
    let mut yaml = String::from("max_queue_size: 1000000\n");
    if threads == 0 {
        yaml.push_str("executors:\n  - { name: \"host\", type: \"DelegatingExecutor\" }\n");
    } else {
        yaml.push_str(&format!(
            "executors:\n  - {{ name: \"cpu\", type: \"ThreadPoolExecutor\", num_threads: {threads} }}\n"
        ));
    }
    yaml.push_str("nodes:\n");
    let executor = if threads == 0 { "host" } else { "cpu" };
    for index in 0..depth {
        let input = if index == 0 {
            "in".to_string()
        } else {
            format!("edge{}", index - 1)
        };
        let output = if index + 1 == depth {
            "out".to_string()
        } else {
            format!("edge{index}")
        };
        yaml.push_str(&format!(
            "  - {{ name: \"node{index}\", kernel: \"BenchPassThrough\", executor: \"{executor}\", input_ports: [\"{input}\"], output_ports: [\"{output}\"] }}\n"
        ));
    }
    yaml.push_str("input_ports: [\"in\"]\noutput_ports: [\"out\"]\n");
    yaml
}

fn bench_round_trip(c: &mut Criterion) {
    common::register_bench_kernels();
    let mut group = c.benchmark_group("rust_end_to_end");
    group.throughput(Throughput::Elements(BATCH));

    for depth in [1usize, 4, 16] {
        let graph = Graph::from_yaml(&chain_yaml(depth, 0)).unwrap();
        let poller = graph.add_poller("out").unwrap();
        graph.start().unwrap();
        let input = graph.input("in").unwrap();
        let timestamp = Cell::new(0i64);

        group.bench_with_input(
            BenchmarkId::new("delegating_i64_depth", depth),
            &depth,
            |benchmark, _| {
                benchmark.iter(|| {
                    for _ in 0..BATCH {
                        let current = timestamp.get();
                        timestamp.set(current + 1);
                        input
                            .send(Packet::from_i64(0).at(Timestamp(current)))
                            .unwrap();
                        black_box(poller.next().unwrap());
                    }
                });
            },
        );
    }
    group.finish();
}

fn bench_thread_pool(c: &mut Criterion) {
    common::register_bench_kernels();
    let mut group = c.benchmark_group("rust_end_to_end");
    group.throughput(Throughput::Elements(BATCH));

    for depth in [1usize, 4, 16] {
        let graph = Graph::from_yaml(&chain_yaml(depth, 4)).unwrap();
        let poller = graph.add_poller("out").unwrap();
        graph.start().unwrap();
        let input = graph.input("in").unwrap();
        let timestamp = Cell::new(0i64);

        group.bench_with_input(
            BenchmarkId::new("pool4_i64_depth", depth),
            &depth,
            |benchmark, _| {
                benchmark.iter(|| {
                    for _ in 0..BATCH {
                        let current = timestamp.get();
                        timestamp.set(current + 1);
                        input
                            .send(Packet::from_i64(0).at(Timestamp(current)))
                            .unwrap();
                    }
                    graph
                        .wait_until_idle_timeout(Duration::from_secs(30))
                        .unwrap();
                    for _ in 0..BATCH {
                        black_box(poller.next().unwrap());
                    }
                });
            },
        );
    }
    group.finish();
}

fn bench_payload_independence(c: &mut Criterion) {
    common::register_bench_kernels();
    let mut group = c.benchmark_group("rust_zero_copy");
    group.throughput(Throughput::Elements(BATCH));

    let large = Packet::from_builtin(Builtin::Buffer(
        BufferData::new(&[512, 512, 3], DT_U8).unwrap(),
    ));
    for (name, template) in [("i64", Packet::from_i64(0)), ("buffer_768k", large)] {
        let graph = Graph::from_yaml(&chain_yaml(1, 0)).unwrap();
        let poller = graph.add_poller("out").unwrap();
        graph.start().unwrap();
        let input = graph.input("in").unwrap();
        let timestamp = Cell::new(0i64);

        group.bench_function(name, |benchmark| {
            benchmark.iter(|| {
                for _ in 0..BATCH {
                    let current = timestamp.get();
                    timestamp.set(current + 1);
                    input.send(template.clone().at(Timestamp(current))).unwrap();
                    black_box(poller.next().unwrap());
                }
            });
        });
    }
    group.finish();
}

criterion_group!(
    benches,
    bench_round_trip,
    bench_thread_pool,
    bench_payload_independence
);
criterion_main!(benches);