request_processing.rs 20.3 KB
Newer Older
1
use criterion::{black_box, criterion_group, criterion_main, BenchmarkId, Criterion, Throughput};
2
use serde_json::{from_str, to_string, to_value, to_vec};
3
4
use std::time::Instant;

5
use sglang_router_rs::core::{BasicWorker, WorkerType};
6
7
8
9
use sglang_router_rs::openai_api_types::{
    ChatCompletionRequest, ChatMessage, CompletionRequest, GenerateParameters, GenerateRequest,
    SamplingParams, StringOrArray, UserMessageContent,
};
10
11
12
13
14
15
16
17
18
19
use sglang_router_rs::routers::bootstrap_injector::inject_bootstrap_fields;

fn create_test_worker() -> BasicWorker {
    BasicWorker::new(
        "http://test-server:8000".to_string(),
        WorkerType::Prefill {
            bootstrap_port: Some(5678),
        },
    )
}
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
/// Create a default GenerateRequest for benchmarks with minimal fields set
fn default_generate_request() -> GenerateRequest {
    GenerateRequest {
        text: None,
        prompt: None,
        input_ids: None,
        stream: false,
        parameters: None,
        sampling_params: None,
        return_logprob: false,
        // SGLang Extensions
        lora_path: None,
        session_params: None,
        return_hidden_states: false,
        rid: None,
    }
}

/// Create a default ChatCompletionRequest for benchmarks with minimal fields set
fn default_chat_completion_request() -> ChatCompletionRequest {
    ChatCompletionRequest {
        model: String::new(),
        messages: vec![],
        max_tokens: None,
        max_completion_tokens: None,
        temperature: None,
        top_p: None,
        n: None,
        stream: false,
        stream_options: None,
        stop: None,
        presence_penalty: None,
        frequency_penalty: None,
        logit_bias: None,
        logprobs: false,
        top_logprobs: None,
        user: None,
        response_format: None,
        seed: None,
        tools: None,
        tool_choice: None,
        parallel_tool_calls: None,
        function_call: None,
        functions: None,
        // SGLang Extensions
        top_k: None,
        min_p: None,
        min_tokens: None,
        repetition_penalty: None,
        regex: None,
        ebnf: None,
        stop_token_ids: None,
        no_stop_trim: false,
        ignore_eos: false,
        continue_final_message: false,
        skip_special_tokens: true,
        // SGLang Extensions
        lora_path: None,
        session_params: None,
        separate_reasoning: true,
        stream_reasoning: true,
        return_hidden_states: false,
    }
}

/// Create a default CompletionRequest for benchmarks with minimal fields set
fn default_completion_request() -> CompletionRequest {
    CompletionRequest {
        model: String::new(),
        prompt: StringOrArray::String(String::new()),
        suffix: None,
        max_tokens: None,
        temperature: None,
        top_p: None,
        n: None,
        stream: false,
        stream_options: None,
        logprobs: None,
        echo: false,
        stop: None,
        presence_penalty: None,
        frequency_penalty: None,
        best_of: None,
        logit_bias: None,
        user: None,
        seed: None,
        // SGLang Extensions
        top_k: None,
        min_p: None,
        min_tokens: None,
        repetition_penalty: None,
        regex: None,
        ebnf: None,
        json_schema: None,
        stop_token_ids: None,
        no_stop_trim: false,
        ignore_eos: false,
        skip_special_tokens: true,
        // SGLang Extensions
        lora_path: None,
        session_params: None,
        return_hidden_states: false,
        other: serde_json::Map::new(),
    }
}

127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
// Sample request data for benchmarks
fn create_sample_generate_request() -> GenerateRequest {
    GenerateRequest {
        text: Some("Write a story about artificial intelligence".to_string()),
        parameters: Some(GenerateParameters {
            max_new_tokens: Some(100),
            temperature: Some(0.8),
            top_p: Some(0.9),
            top_k: Some(50),
            repetition_penalty: Some(1.0),
            ..Default::default()
        }),
        sampling_params: Some(SamplingParams {
            temperature: Some(0.8),
            top_p: Some(0.9),
            top_k: Some(50),
            frequency_penalty: Some(0.0),
            presence_penalty: Some(0.0),
            repetition_penalty: Some(1.0),
            ..Default::default()
        }),
148
        ..default_generate_request()
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
    }
}

fn create_sample_chat_completion_request() -> ChatCompletionRequest {
    ChatCompletionRequest {
        model: "gpt-3.5-turbo".to_string(),
        messages: vec![
            ChatMessage::System {
                role: "system".to_string(),
                content: "You are a helpful assistant".to_string(),
                name: None,
            },
            ChatMessage::User {
                role: "user".to_string(),
                content: UserMessageContent::Text(
                    "Explain quantum computing in simple terms".to_string(),
                ),
                name: None,
            },
        ],
        max_tokens: Some(150),
        max_completion_tokens: Some(150),
        temperature: Some(0.7),
        top_p: Some(1.0),
        n: Some(1),
        presence_penalty: Some(0.0),
        frequency_penalty: Some(0.0),
        parallel_tool_calls: Some(true),
177
        ..default_chat_completion_request()
178
179
180
181
182
183
184
185
186
187
188
189
190
191
    }
}

fn create_sample_completion_request() -> CompletionRequest {
    CompletionRequest {
        model: "text-davinci-003".to_string(),
        prompt: StringOrArray::String("Complete this sentence: The future of AI is".to_string()),
        max_tokens: Some(50),
        temperature: Some(0.8),
        top_p: Some(1.0),
        n: Some(1),
        presence_penalty: Some(0.0),
        frequency_penalty: Some(0.0),
        best_of: Some(1),
192
        ..default_completion_request()
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
    }
}

fn create_large_chat_completion_request() -> ChatCompletionRequest {
    let mut messages = vec![ChatMessage::System {
        role: "system".to_string(),
        content: "You are a helpful assistant with extensive knowledge.".to_string(),
        name: None,
    }];

    // Add many user/assistant pairs to simulate a long conversation
    for i in 0..50 {
        messages.push(ChatMessage::User {
            role: "user".to_string(),
            content: UserMessageContent::Text(format!("Question {}: What do you think about topic number {} which involves complex reasoning about multiple interconnected systems and their relationships?", i, i)),
            name: None,
        });
        messages.push(ChatMessage::Assistant {
            role: "assistant".to_string(),
            content: Some(format!("Answer {}: This is a detailed response about topic {} that covers multiple aspects and provides comprehensive analysis of the interconnected systems you mentioned.", i, i)),
            name: None,
            tool_calls: None,
            function_call: None,
216
            reasoning_content: None,
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
        });
    }

    ChatCompletionRequest {
        model: "gpt-4".to_string(),
        messages,
        max_tokens: Some(1000),
        max_completion_tokens: Some(1000),
        temperature: Some(0.7),
        top_p: Some(0.95),
        n: Some(1),
        presence_penalty: Some(0.1),
        frequency_penalty: Some(0.1),
        top_logprobs: Some(5),
        user: Some("benchmark_user".to_string()),
        seed: Some(42),
        parallel_tool_calls: Some(true),
234
        ..default_chat_completion_request()
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
    }
}

// Benchmark JSON serialization
fn bench_json_serialization(c: &mut Criterion) {
    let mut group = c.benchmark_group("json_serialization");

    let generate_req = create_sample_generate_request();
    let chat_req = create_sample_chat_completion_request();
    let completion_req = create_sample_completion_request();
    let large_chat_req = create_large_chat_completion_request();

    group.bench_function("generate_request", |b| {
        b.iter(|| {
            let json = to_string(black_box(&generate_req)).unwrap();
            black_box(json);
        });
    });

    group.bench_function("chat_completion_request", |b| {
        b.iter(|| {
            let json = to_string(black_box(&chat_req)).unwrap();
            black_box(json);
        });
    });

    group.bench_function("completion_request", |b| {
        b.iter(|| {
            let json = to_string(black_box(&completion_req)).unwrap();
            black_box(json);
        });
    });

    group.bench_function("large_chat_completion_request", |b| {
        b.iter(|| {
            let json = to_string(black_box(&large_chat_req)).unwrap();
            black_box(json);
        });
    });

    group.bench_function("generate_request_to_bytes", |b| {
        b.iter(|| {
            let bytes = to_vec(black_box(&generate_req)).unwrap();
            black_box(bytes);
        });
    });

    group.finish();
}

// Benchmark JSON deserialization
fn bench_json_deserialization(c: &mut Criterion) {
    let mut group = c.benchmark_group("json_deserialization");

    let generate_json = to_string(&create_sample_generate_request()).unwrap();
    let chat_json = to_string(&create_sample_chat_completion_request()).unwrap();
    let completion_json = to_string(&create_sample_completion_request()).unwrap();
    let large_chat_json = to_string(&create_large_chat_completion_request()).unwrap();

    group.bench_function("generate_request", |b| {
        b.iter(|| {
            let req: GenerateRequest = from_str(black_box(&generate_json)).unwrap();
            black_box(req);
        });
    });

    group.bench_function("chat_completion_request", |b| {
        b.iter(|| {
            let req: ChatCompletionRequest = from_str(black_box(&chat_json)).unwrap();
            black_box(req);
        });
    });

    group.bench_function("completion_request", |b| {
        b.iter(|| {
            let req: CompletionRequest = from_str(black_box(&completion_json)).unwrap();
            black_box(req);
        });
    });

    group.bench_function("large_chat_completion_request", |b| {
        b.iter(|| {
            let req: ChatCompletionRequest = from_str(black_box(&large_chat_json)).unwrap();
            black_box(req);
        });
    });

    group.finish();
}

325
326
327
// Benchmark bootstrap injection (replaces request adaptation)
fn bench_bootstrap_injection(c: &mut Criterion) {
    let mut group = c.benchmark_group("bootstrap_injection");
328
329
330
331
332

    let generate_req = create_sample_generate_request();
    let chat_req = create_sample_chat_completion_request();
    let completion_req = create_sample_completion_request();
    let large_chat_req = create_large_chat_completion_request();
333
    let worker = create_test_worker();
334

335
    group.bench_function("generate_bootstrap_injection", |b| {
336
        b.iter(|| {
337
338
339
            let mut json = to_value(black_box(&generate_req)).unwrap();
            inject_bootstrap_fields(&mut json, &worker).unwrap();
            black_box(json);
340
341
342
        });
    });

343
    group.bench_function("chat_completion_bootstrap_injection", |b| {
344
        b.iter(|| {
345
346
347
            let mut json = to_value(black_box(&chat_req)).unwrap();
            inject_bootstrap_fields(&mut json, &worker).unwrap();
            black_box(json);
348
349
350
        });
    });

351
    group.bench_function("completion_bootstrap_injection", |b| {
352
        b.iter(|| {
353
354
355
            let mut json = to_value(black_box(&completion_req)).unwrap();
            inject_bootstrap_fields(&mut json, &worker).unwrap();
            black_box(json);
356
357
358
        });
    });

359
    group.bench_function("large_chat_completion_bootstrap_injection", |b| {
360
        b.iter(|| {
361
362
363
            let mut json = to_value(black_box(&large_chat_req)).unwrap();
            inject_bootstrap_fields(&mut json, &worker).unwrap();
            black_box(json);
364
365
366
367
368
369
        });
    });

    group.finish();
}

370
371
372
// Benchmark direct JSON routing (replaces regular routing)
fn bench_direct_json_routing(c: &mut Criterion) {
    let mut group = c.benchmark_group("direct_json_routing");
373
374
375
376
377
378
379

    let generate_req = create_sample_generate_request();
    let chat_req = create_sample_chat_completion_request();
    let completion_req = create_sample_completion_request();

    group.bench_function("generate_to_json", |b| {
        b.iter(|| {
380
381
382
383
384
385
386
387
            let json = to_value(black_box(&generate_req)).unwrap();
            black_box(json);
        });
    });

    group.bench_function("generate_to_json_string", |b| {
        b.iter(|| {
            let json = to_string(black_box(&generate_req)).unwrap();
388
389
390
391
392
393
            black_box(json);
        });
    });

    group.bench_function("generate_to_bytes", |b| {
        b.iter(|| {
394
            let bytes = to_vec(black_box(&generate_req)).unwrap();
395
396
397
398
399
400
            black_box(bytes);
        });
    });

    group.bench_function("chat_completion_to_json", |b| {
        b.iter(|| {
401
            let json = to_value(black_box(&chat_req)).unwrap();
402
403
404
405
            black_box(json);
        });
    });

406
    group.bench_function("chat_completion_to_json_string", |b| {
407
        b.iter(|| {
408
409
            let json = to_string(black_box(&chat_req)).unwrap();
            black_box(json);
410
411
412
413
414
        });
    });

    group.bench_function("completion_to_json", |b| {
        b.iter(|| {
415
            let json = to_value(black_box(&completion_req)).unwrap();
416
417
418
419
420
421
422
423
424
425
426
427
428
429
            black_box(json);
        });
    });

    group.finish();
}

// Benchmark throughput with different request sizes
fn bench_throughput_by_size(c: &mut Criterion) {
    let mut group = c.benchmark_group("throughput_by_size");

    // Create requests of different sizes
    let small_generate = GenerateRequest {
        text: Some("Hi".to_string()),
430
        ..default_generate_request()
431
432
433
434
    };

    let medium_generate = GenerateRequest {
        text: Some("Write a medium length story about AI".repeat(10)),
435
        ..default_generate_request()
436
437
438
439
    };

    let large_generate = GenerateRequest {
        text: Some("Write a very long and detailed story about artificial intelligence and its impact on society".repeat(100)),
440
        ..default_generate_request()
441
442
    };

443
444
    let worker = create_test_worker();

445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
    for (name, req) in [
        ("small", &small_generate),
        ("medium", &medium_generate),
        ("large", &large_generate),
    ] {
        let json = to_string(req).unwrap();
        let size_bytes = json.len();

        group.throughput(Throughput::Bytes(size_bytes as u64));
        group.bench_with_input(BenchmarkId::new("serialize", name), &req, |b, req| {
            b.iter(|| {
                let json = to_string(black_box(req)).unwrap();
                black_box(json);
            });
        });

        group.bench_with_input(
            BenchmarkId::new("deserialize", name),
            &json,
            |b, json_str| {
                b.iter(|| {
                    let req: GenerateRequest = black_box(from_str(json_str)).unwrap();
                    black_box(req);
                });
            },
        );

472
473
474
475
476
477
478
479
480
481
482
        group.bench_with_input(
            BenchmarkId::new("bootstrap_inject", name),
            &req,
            |b, req| {
                b.iter(|| {
                    let mut json = to_value(req).unwrap();
                    inject_bootstrap_fields(&mut json, &worker).unwrap();
                    black_box(json);
                });
            },
        );
483
484
485
486
487
    }

    group.finish();
}

488
// Benchmark full round-trip: deserialize -> inject bootstrap -> serialize
489
490
491
492
493
494
fn bench_full_round_trip(c: &mut Criterion) {
    let mut group = c.benchmark_group("full_round_trip");

    let generate_json = to_string(&create_sample_generate_request()).unwrap();
    let chat_json = to_string(&create_sample_chat_completion_request()).unwrap();
    let completion_json = to_string(&create_sample_completion_request()).unwrap();
495
    let worker = create_test_worker();
496
497
498
499
500

    group.bench_function("generate_openai_to_pd_pipeline", |b| {
        b.iter(|| {
            // Deserialize OpenAI request
            let req: GenerateRequest = from_str(black_box(&generate_json)).unwrap();
501
502
503
504
505
506
            // Convert to JSON Value
            let mut json = to_value(&req).unwrap();
            // Inject bootstrap fields
            inject_bootstrap_fields(&mut json, &worker).unwrap();
            // Serialize final request
            let pd_json = to_string(&json).unwrap();
507
508
509
510
511
512
513
            black_box(pd_json);
        });
    });

    group.bench_function("chat_completion_openai_to_pd_pipeline", |b| {
        b.iter(|| {
            let req: ChatCompletionRequest = from_str(black_box(&chat_json)).unwrap();
514
515
516
            let mut json = to_value(&req).unwrap();
            inject_bootstrap_fields(&mut json, &worker).unwrap();
            let pd_json = to_string(&json).unwrap();
517
518
519
520
521
522
523
            black_box(pd_json);
        });
    });

    group.bench_function("completion_openai_to_pd_pipeline", |b| {
        b.iter(|| {
            let req: CompletionRequest = from_str(black_box(&completion_json)).unwrap();
524
525
526
            let mut json = to_value(&req).unwrap();
            inject_bootstrap_fields(&mut json, &worker).unwrap();
            let pd_json = to_string(&json).unwrap();
527
528
529
530
            black_box(pd_json);
        });
    });

531
    group.bench_function("generate_direct_json_pipeline", |b| {
532
533
534
        b.iter(|| {
            // Deserialize OpenAI request
            let req: GenerateRequest = from_str(black_box(&generate_json)).unwrap();
535
536
537
538
            // Convert to JSON for direct routing (no bootstrap injection)
            let routing_json = to_value(&req).unwrap();
            let json_string = to_string(&routing_json).unwrap();
            black_box(json_string);
539
540
541
542
543
544
545
546
547
548
549
550
551
552
        });
    });

    group.finish();
}

fn benchmark_summary(c: &mut Criterion) {
    let group = c.benchmark_group("benchmark_summary");

    println!("\nSGLang Router Performance Benchmark Suite");
    println!("=============================================");

    // Quick performance overview
    let generate_req = create_sample_generate_request();
553
    let worker = create_test_worker();
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576

    println!("\nQuick Performance Overview:");

    // Measure serialization
    let start = Instant::now();
    for _ in 0..1000 {
        let _ = black_box(to_string(&generate_req).unwrap());
    }
    let serialize_time = start.elapsed().as_nanos() / 1000;
    println!("  * Serialization (avg):     {:>8} ns/req", serialize_time);

    // Measure deserialization
    let json = to_string(&generate_req).unwrap();
    let start = Instant::now();
    for _ in 0..1000 {
        let _: GenerateRequest = black_box(from_str(&json).unwrap());
    }
    let deserialize_time = start.elapsed().as_nanos() / 1000;
    println!(
        "  * Deserialization (avg):   {:>8} ns/req",
        deserialize_time
    );

577
    // Measure bootstrap injection (replaces adaptation)
578
579
    let start = Instant::now();
    for _ in 0..1000 {
580
581
        let mut json = to_value(&generate_req).unwrap();
        let _ = black_box(inject_bootstrap_fields(&mut json, &worker));
582
    }
583
584
    let inject_time = start.elapsed().as_nanos() / 1000;
    println!("  * Bootstrap Injection (avg): {:>6} ns/req", inject_time);
585
586

    // Calculate ratios
587
    let total_pipeline = serialize_time + deserialize_time + inject_time;
588
589
590
591
592
593
    println!("  * Total Pipeline (avg):    {:>8} ns/req", total_pipeline);

    println!("\nPerformance Insights:");
    if deserialize_time > serialize_time * 2 {
        println!("  • Deserialization is significantly faster than serialization");
    }
594
    if inject_time < serialize_time / 10 {
595
        println!(
596
597
            "  • Bootstrap injection overhead is negligible ({:.1}% of serialization)",
            (inject_time as f64 / serialize_time as f64) * 100.0
598
599
        );
    }
600
601
    if total_pipeline < 100_000 {
        println!("  • Total pipeline latency is excellent (< 100μs)");
602
603
    }

604
605
606
607
608
609
    println!("\nSimplification Benefits:");
    println!("  • Eliminated complex type conversion layer");
    println!("  • Reduced memory allocations");
    println!("  • Automatic field preservation (no manual mapping)");
    println!("  • Direct JSON manipulation improves performance");

610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
    println!("\nRecommendations:");
    if serialize_time > deserialize_time {
        println!("  • Focus optimization efforts on serialization rather than deserialization");
    }
    println!("  • PD mode overhead is minimal - safe to use for latency-sensitive workloads");
    println!("  • Consider batching small requests to improve overall throughput");

    println!("\n{}", "=".repeat(50));

    group.finish();
}

criterion_group!(
    benches,
    benchmark_summary,
    bench_json_serialization,
    bench_json_deserialization,
627
628
    bench_bootstrap_injection,
    bench_direct_json_routing,
629
630
631
632
    bench_throughput_by_size,
    bench_full_round_trip
);
criterion_main!(benches);