2nd/10_Wiki/Topics/AI_and_ML/Bottlenecks.md

---
id: wiki-2026-0508-bottlenecks
title: Bottlenecks (Performance & Process)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [병목, bottleneck, theory of constraints, TOC, critical path, profiling]
duplicate_of: none
source_trust_level: A
confidence_score: 0.93
verification_status: applied
tags: [performance, bottleneck, profiling, theory-of-constraints, optimization, scalability, latency]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: any
  framework: profiling tools
---

# Bottlenecks

## 📌 한 줄 통찰
> **"매 system 의 throat"**. 매 weakest link 의 throughput 의 결정. 매 non-bottleneck 의 improve = 매 시간 낭비. 매 Goldratt's TOC: 매 5 step. 매 modern AI: 매 HBM bandwidth + 매 network 의 bottleneck.

## 📖 핵심

### 매 type
1. **Hardware**: CPU / GPU / RAM / disk / network.
2. **Software**: algorithm / blocking / lock contention.
3. **Process**: approval / single point of expertise.
4. **Data**: schema / indexing / partitioning.
5. **Cognitive** (team): meeting / context-switch.

### Theory of Constraints (Goldratt)
1. **Identify** the bottleneck.
2. **Exploit** it (use 100%).
3. **Subordinate** non-bottleneck (don't over-feed).
4. **Elevate** it (invest to widen).
5. **Repeat** (new bottleneck emerges).

### Amdahl's Law (related)
- 매 90% 의 100× → 매 전체 의 매 10× cap.
- 매 bottleneck 의 X 의 의미.

### 매 hardware bottleneck 의 modern (LLM)
- **HBM bandwidth**: 매 H100 = 매 3 TB/s. 매 LLM inference 의 dominant.
- **NVLink**: 매 GPU-GPU.
- **Network** (RDMA, InfiniBand): 매 distributed train.
- **PCIe**: 매 GPU-CPU.
- **Storage**: 매 NVMe vs spinning.
- **Power / cooling**: 매 datacenter limit.

### 매 software bottleneck
- **CPU-bound**: 매 compute heavy.
- **I/O-bound**: 매 disk / network wait.
- **Memory-bound**: 매 swap / cache miss.
- **Lock contention**: 매 mutex.
- **GIL** (Python): 매 single-thread CPU.
- **N+1 query**: 매 ORM 의 typical.

### 매 detection
- **Profiler**: cProfile, perf, async-profiler.
- **Trace**: distributed tracing (Jaeger).
- **Metric**: CPU/mem/disk/network util.
- **APM**: Datadog, NewRelic.
- **Flame graph**.
- **Critical path**.

### 매 process bottleneck
- 매 approval chain.
- 매 single expert.
- 매 environment provisioning.
- 매 review SLA.
- 매 meeting cadence.

→ 매 DORA Lead Time 의 component.

### 매 data bottleneck
- 매 single hot row.
- 매 missing index.
- 매 cross-shard transaction.
- 매 schema migration block.

### 매 distributed bottleneck (modern)
- 매 leader 의 single (Raft, Paxos).
- 매 cross-region call.
- 매 sync replication.
- 매 connection pool limit.

## 💻 패턴

### Profile (Python cProfile)
```python
import cProfile, pstats

def main():
    expensive_call()
    cheap_call()

cProfile.run('main()', 'out.prof')
stats = pstats.Stats('out.prof').sort_stats('cumulative')
stats.print_stats(20)
```

### Linux perf (system-level)
```bash
# 매 CPU profile
perf record -F 99 -p $PID -- sleep 10
perf report

# 매 flame graph
perf script | ./stackcollapse-perf.pl | ./flamegraph.pl > flame.svg
```

### Async profiler (JVM)
```bash
# 매 sample lock contention
java -jar async-profiler.jar -e lock -d 30 -f lock.html $PID

# 매 wall clock (I/O bound 도)
java -jar async-profiler.jar -e wall -d 30 -f wall.html $PID
```

### N+1 detect (Django)
```python
from django.test.utils import CaptureQueriesContext
from django.db import connection

with CaptureQueriesContext(connection) as ctx:
    posts = Post.objects.all()
    for post in posts:
        print(post.author.name)  # 매 N+1
    
    if len(ctx.captured_queries) > 5:
        log(f'N+1 detected: {len(ctx.captured_queries)} queries')

# 매 fix
posts = Post.objects.select_related('author')  # 매 1 query
```

### GPU bottleneck profile (PyTorch)
```python
import torch.profiler as prof

with prof.profile(
    activities=[prof.ProfilerActivity.CPU, prof.ProfilerActivity.CUDA],
    record_shapes=True,
    profile_memory=True,
) as p:
    model(input)

print(p.key_averages().table(sort_by='cuda_time_total', row_limit=20))

# 매 HBM bandwidth bottleneck 의 reveal
```

### Lock contention detection
```python
import threading

class LockMonitor:
    def __init__(self, lock):
        self.lock = lock
        self.wait_times = []
    
    def __enter__(self):
        start = time.time()
        self.lock.acquire()
        self.wait_times.append(time.time() - start)
    
    def __exit__(self, *args):
        self.lock.release()
    
    def report(self):
        if not self.wait_times: return
        avg = sum(self.wait_times) / len(self.wait_times)
        if avg > 0.1: log(f'Lock contention: avg wait {avg*1000}ms')
```

### Distributed trace (Jaeger)
```python
from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.exporter.jaeger.thrift import JaegerExporter

trace.set_tracer_provider(TracerProvider())
tracer = trace.get_tracer(__name__)

@tracer.start_as_current_span('handle_request')
def handle(req):
    with tracer.start_as_current_span('db_query') as span:
        span.set_attribute('db.statement', 'SELECT ...')
        result = db.query(...)
    return result
```

→ 매 시각적 bottleneck identify.

### Process bottleneck (workflow analysis)
```python
def analyze_workflow(stage_durations):
    """매 stage 별 의 throughput 의 비교."""
    rates = {stage: 1 / dur for stage, dur in stage_durations.items()}
    bottleneck = min(rates, key=rates.get)
    
    overall_rate = rates[bottleneck]
    waste = sum(r - overall_rate for r in rates.values() if r > overall_rate)
    
    return {
        'bottleneck': bottleneck,
        'overall_rate_per_min': overall_rate * 60,
        'capacity_wasted': waste,
    }
```

### Critical path (DAG)
```python
import networkx as nx

def critical_path(tasks):
    """매 longest path through DAG."""
    G = nx.DiGraph()
    for task in tasks:
        G.add_node(task.id, duration=task.duration)
        for dep in task.deps:
            G.add_edge(dep, task.id)
    
    # 매 longest path
    return nx.dag_longest_path(G, weight='duration')
```

## 🤔 결정 기준
| 증상 | Tool |
|---|---|
| Slow request | APM + distributed trace |
| CPU pegged | Flame graph (perf) |
| GPU underutilized | Memory bandwidth (PyTorch profiler) |
| Slow query | EXPLAIN + slow query log |
| Lock contention | async-profiler -e lock |
| Long lead time | Process / DORA analysis |
| Thundering herd | Coordination check |

**기본값**: 매 measure first. 매 hypothesis-based optimize.

## 🔗 Graph
- 부모: [[System-Design]]
- 변형: [[CPU-Bound]]
- 응용: [[Theory-of-Constraints]] · [[Amdahls Law (암달의 법칙)]] · [[Critical-Path]]
- Tool: [[Profiling]] · [[Flame-Graph]] · [[Distributed-Tracing]]
- Adjacent: [[Optimization]] · [[Scalability]] · [[DORA-Metrics]]

## 🤖 LLM 활용
**언제**: 매 performance optimization. 매 capacity planning. 매 incident root cause. 매 process improvement.
**언제 X**: 매 hypothesis 없 의 optimize.

## ❌ 안티패턴
- **Optimize without measure**: 매 wrong place.
- **Non-bottleneck improve**: 매 시간 waste (TOC).
- **모든 part 의 평등 invest**: 매 ROI low.
- **Single profile 의 trust**: 매 representative X.
- **Process 의 "사람 의 fault"**: 매 system issue 가 대부분.
- **Premature optimization**: 매 simplicity lose.

## 🧪 검증 / 중복
- Verified (Goldratt TOC, Knuth premature optimization, Brendan Gregg Systems Performance).
- 신뢰도 A.
- Related: [[Amdahls Law (암달의 법칙)]] · [[Theory-of-Constraints]] · [[Profiling]] · [[Critical-Path]].

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — type + TOC + 매 profile / N+1 / GPU / trace code |