File size: 12,396 Bytes
e706de2 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 |
# Concept: Parallel Processing & Performance Optimization
## Overview
This example demonstrates **concurrent execution** of multiple LLM requests using separate context sequences, a critical technique for building scalable AI agent systems.
## The Performance Problem
### Sequential Processing (Slow)
Traditional approach processes one request at a time:
```
Request 1 ────────→ Response 1 (2s)
↓
Request 2 ────────→ Response 2 (2s)
↓
Total: 4 seconds
```
### Parallel Processing (Fast)
This example processes multiple requests simultaneously:
```
Request 1 ────────→ Response 1 (2s) ──┐
├→ Total: 2 seconds
Request 2 ────────→ Response 2 (2s) ──┘
(Both running at the same time)
```
**Performance gain: 2x speedup!**
## Core Concept: Context Sequences
### Single vs. Multiple Sequences
```
┌────────────────────────────────────────────────┐
│ Model (Loaded Once) │
├────────────────────────────────────────────────┤
│ Context │
│ ┌──────────────┐ ┌──────────────┐ │
│ │ Sequence 1 │ │ Sequence 2 │ │
│ │ │ │ │ │
│ │ Conversation │ │ Conversation │ │
│ │ History A │ │ History B │ │
│ └──────────────┘ └──────────────┘ │
└────────────────────────────────────────────────┘
```
**Key insights:**
- Model weights are shared (memory efficient)
- Each sequence has independent history
- Sequences can process in parallel
- Both use the same underlying model
## How Parallel Processing Works
### Promise.all Pattern
JavaScript's `Promise.all()` enables concurrent execution:
```
Sequential:
────────────────────────────────────
await fn1(); // Wait 2s
await fn2(); // Wait 2s more
Total: 4s
Parallel:
────────────────────────────────────
await Promise.all([
fn1(), // Start immediately
fn2() // Start immediately (don't wait!)
]);
Total: 2s (whichever finishes last)
```
### Execution Timeline
```
Time → 0s 1s 2s 3s 4s
│ │ │ │ │
Seq 1: ├───────Processing───────┤
│ └─ Response 1
│
Seq 2: ├───────Processing───────┤
└─ Response 2
Both complete at ~2s instead of 4s!
```
## GPU Batch Processing
### Why Batching Matters
Modern GPUs process multiple operations efficiently:
```
Without Batching (Inefficient)
──────────────────────────────
GPU: [Token 1] ... wait ...
GPU: [Token 2] ... wait ...
GPU: [Token 3] ... wait ...
└─ GPU underutilized
With Batching (Efficient)
─────────────────────────
GPU: [Tokens 1-1024] ← Full batch
└─ GPU fully utilized!
```
**batchSize parameter**: Controls how many tokens process together.
### Trade-offs
```
Small Batch (e.g., 128) Large Batch (e.g., 2048)
─────────────────────── ────────────────────────
✓ Lower memory ✓ Better GPU utilization
✓ More flexible ✓ Faster throughput
✗ Slower throughput ✗ Higher memory usage
✗ GPU underutilized ✗ May exceed VRAM
```
**Sweet spot**: Usually 512-1024 for consumer GPUs.
## Architecture Patterns
### Pattern 1: Multi-User Service
```
┌─────────┐ ┌─────────┐ ┌─────────┐
│ User A │ │ User B │ │ User C │
└────┬────┘ └────┬────┘ └────┬────┘
│ │ │
└────────────┼────────────┘
↓
┌────────────────┐
│ Load Balancer │
└────────────────┘
↓
┌────────────┼────────────┐
↓ ↓ ↓
┌─────────┐ ┌─────────┐ ┌─────────┐
│ Seq 1 │ │ Seq 2 │ │ Seq 3 │
└─────────┘ └─────────┘ └─────────┘
└────────────┼────────────┘
↓
┌────────────────┐
│ Shared Model │
└────────────────┘
```
### Pattern 2: Multi-Agent System
```
┌──────────────┐
│ Task │
└──────┬───────┘
│
┌────────┼────────┐
↓ ↓ ↓
┌────────┐ ┌──────┐ ┌──────────┐
│Planner │ │Critic│ │ Executor │
│ Agent │ │Agent │ │ Agent │
└───┬────┘ └──┬───┘ └────┬─────┘
│ │ │
└─────────┼──────────┘
↓
(All run in parallel)
```
### Pattern 3: Pipeline Processing
```
Input Queue: [Task1, Task2, Task3, ...]
↓
┌───────────────┐
│ Dispatcher │
└───────────────┘
↓
┌───────────┼───────────┐
↓ ↓ ↓
Sequence 1 Sequence 2 Sequence 3
↓ ↓ ↓
└───────────┼───────────┘
↓
Output: [R1, R2, R3]
```
## Resource Management
### Memory Allocation
Each sequence consumes memory:
```
┌──────────────────────────────────┐
│ Total VRAM: 8GB │
├──────────────────────────────────┤
│ Model Weights: 4.0 GB │
│ Context Base: 1.0 GB │
│ Sequence 1 (KV Cache): 0.8 GB │
│ Sequence 2 (KV Cache): 0.8 GB │
│ Sequence 3 (KV Cache): 0.8 GB │
│ Overhead: 0.6 GB │
├──────────────────────────────────┤
│ Total Used: 8.0 GB │
│ Remaining: 0.0 GB │
└──────────────────────────────────┘
Maximum capacity!
```
**Formula**:
```
Required VRAM = Model + Context + (NumSequences × KVCache)
```
### Finding Optimal Sequence Count
```
Too Few (1-2) Optimal (4-8) Too Many (16+)
───────────── ───────────── ──────────────
GPU underutilized Balanced use Memory overflow
↓ ↓ ↓
Slow throughput Best performance Thrashing/crashes
```
**Test your system**:
1. Start with 2 sequences
2. Monitor VRAM usage
3. Increase until performance plateaus
4. Back off if memory issues occur
## Real-World Scenarios
### Scenario 1: Chatbot Service
```
Challenge: 100 users, each waiting 2s per response
Sequential: 100 × 2s = 200s (3.3 minutes!)
Parallel (10 seq): 10 batches × 2s = 20s
10x speedup!
```
### Scenario 2: Batch Analysis
```
Task: Analyze 1000 documents
Sequential: 1000 × 3s = 50 minutes
Parallel (8 seq): 125 batches × 3s = 6.25 minutes
8x speedup!
```
### Scenario 3: Multi-Agent Collaboration
```
Agents: Planner, Analyzer, Executor (all needed)
Sequential: Wait for each → Slow pipeline
Parallel: All work together → Fast decision-making
```
## Limitations & Considerations
### 1. Context Capacity Sharing
```
Problem: Sequences share total context space
───────────────────────────────────────────
Total context: 4096 tokens
2 sequences: Each gets ~2048 tokens max
4 sequences: Each gets ~1024 tokens max
More sequences = Less history per sequence!
```
### 2. CPU vs GPU Parallelism
```
With GPU: CPU Only:
True parallel processing Interleaved processing
Multiple CUDA streams Single thread context-switching
(Still helps throughput!)
```
### 3. Not Always Faster
```
When parallel helps: When it doesn't:
• Independent requests • Dependent requests (must wait)
• I/O-bound operations • Very short prompts (overhead)
• Multiple users • Single sequential conversation
```
## Best Practices
### 1. Design for Independence
```
✓ Good: Separate user conversations
✓ Good: Independent analysis tasks
✗ Bad: Sequential reasoning steps (use ReAct instead)
```
### 2. Monitor Resources
```
Track:
• VRAM usage per sequence
• Processing time per request
• Queue depths
• Error rates
```
### 3. Implement Graceful Degradation
```
if (vramExceeded) {
reduceSequenceCount();
// or queue requests instead
}
```
### 4. Handle Errors Properly
```javascript
try {
const results = await Promise.all([...]);
} catch (error) {
// One failure doesn't crash all sequences
handlePartialResults();
}
```
## Comparison: Evolution of Performance
```
Stage Requests/Min Pattern
───────────────── ───────────── ───────────────
1. Basic (intro) 30 Sequential
2. Batch (this) 120 4 sequences
3. Load balanced 240 8 sequences + queue
4. Distributed 1000+ Multiple machines
```
## Key Takeaways
1. **Parallelism is essential** for production AI agent systems
2. **Sequences share model** but maintain independent state
3. **Promise.all** enables concurrent JavaScript execution
4. **Batch size** affects GPU utilization and throughput
5. **Memory is the limit** - more sequences need more VRAM
6. **Not magic** - only helps with independent tasks
## Practical Formula
```
Speedup = min(
Number_of_Sequences,
Available_VRAM / Memory_Per_Sequence,
GPU_Compute_Limit
)
```
Typically: 2-10x speedup for well-designed systems.
This technique is foundational for building scalable agent architectures that can handle real-world workloads efficiently.
|