perf-profiler（性能分析器）

Performance Profiler

Measure, profile, and optimize application performance. Covers CPU profiling, memory analysis, flame graphs, benchmarking, load testing, and language-specific optimization patterns.

When to Use

• Diagnosing why an application or function is slow
• Measuring CPU and memory usage
• Generating flame graphs to visualize hot paths
• Benchmarking functions or endpoints
• Load testing APIs before deployment
• Finding and fixing memory leaks
• Optimizing database query performance
• Comparing performance before and after changes

Quick Timing

Command-line timing

# Time any command
time my-command --flag

# More precise: multiple runs with stats
for i in $(seq 1 10); do
  /usr/bin/time -f "%e" my-command 2>&1
done | awk '{sum+=$1; sumsq+=$1*$1; count++} END {
  avg=sum/count;
  stddev=sqrt(sumsq/count - avg*avg);
  printf "runs=%d avg=%.3fs stddev=%.3fs\
", count, avg, stddev
}'

# Hyperfine (better benchmarking tool)
# Install: https://github.com/sharkdp/hyperfine
hyperfine 'command-a' 'command-b'
hyperfine --warmup 3 --runs 20 'my-command'
hyperfine --export-json results.json 'old-version' 'new-version'

Inline timing (any language)

// Node.js
console.time('operation');
await doExpensiveThing();
console.timeEnd('operation'); // "operation: 142.3ms"

// High-resolution
const start = performance.now();
await doExpensiveThing();
const elapsed = performance.now() - start;
console.log(`Elapsed: ${elapsed.toFixed(2)}ms`);

# Python
import time

start = time.perf_counter()
do_expensive_thing()
elapsed = time.perf_counter() - start
print(f"Elapsed: {elapsed:.4f}s")

# Context manager
from contextlib import contextmanager

@contextmanager
def timer(label=""):
    start = time.perf_counter()
    yield
    elapsed = time.perf_counter() - start
    print(f"{label}: {elapsed:.4f}s")

with timer("data processing"):
    process_data()

// Go
start := time.Now()
doExpensiveThing()
fmt.Printf("Elapsed: %v\
", time.Since(start))

Node.js Profiling

CPU profiling with V8 inspector

# Generate CPU profile (writes .cpuprofile file)
node --cpu-prof app.js
# Open the .cpuprofile in Chrome DevTools > Performance tab

# Profile for a specific duration
node --cpu-prof --cpu-prof-interval=100 app.js

# Inspect running process
node --inspect app.js
# Open chrome://inspect in Chrome, click "inspect"
# Go to Performance tab, click Record

Heap snapshots (memory)

# Generate heap snapshot
node --heap-prof app.js

# Take snapshots programmatically
node -e "
const v8 = require('v8');
const fs = require('fs');

// Take snapshot
const snapshotStream = v8.writeHeapSnapshot();
console.log('Heap snapshot written to:', snapshotStream);
"

# Compare heap snapshots to find leaks:
# 1. Take snapshot A (baseline)
# 2. Run operations that might leak
# 3. Take snapshot B
# 4. In Chrome DevTools > Memory, load both and use "Comparison" view

Memory usage monitoring

// Print memory usage periodically
setInterval(() => {
  const usage = process.memoryUsage();
  console.log({
    rss: `${(usage.rss / 1024 / 1024).toFixed(1)}MB`,
    heapUsed: `${(usage.heapUsed / 1024 / 1024).toFixed(1)}MB`,
    heapTotal: `${(usage.heapTotal / 1024 / 1024).toFixed(1)}MB`,
    external: `${(usage.external / 1024 / 1024).toFixed(1)}MB`,
  });
}, 5000);

// Detect memory growth
let lastHeap = 0;
setInterval(() => {
  const heap = process.memoryUsage().heapUsed;
  const delta = heap - lastHeap;
  if (delta > 1024 * 1024) { // > 1MB growth
    console.warn(`Heap grew by ${(delta / 1024 / 1024).toFixed(1)}MB`);
  }
  lastHeap = heap;
}, 10000);

Node.js benchmarking

// Simple benchmark function
function benchmark(name, fn, iterations = 10000) {
  // Warmup
  for (let i = 0; i < 100; i++) fn();

  const start = performance.now();
  for (let i = 0; i < iterations; i++) fn();
  const elapsed = performance.now() - start;

  console.log(`${name}: ${(elapsed / iterations).toFixed(4)}ms/op (${iterations} iterations in ${elapsed.toFixed(1)}ms)`);
}

benchmark('JSON.parse', () => JSON.parse('{"key":"value","num":42}'));
benchmark('regex match', () => /^\d{4}-\d{2}-\d{2}$/.test('2026-02-03'));

Python Profiling

cProfile (built-in CPU profiler)

# Profile a script
python3 -m cProfile -s cumulative my_script.py

# Save to file for analysis
python3 -m cProfile -o profile.prof my_script.py

# Analyze saved profile
python3 -c "
import pstats
stats = pstats.Stats('profile.prof')
stats.sort_stats('cumulative')
stats.print_stats(20)
"

# Profile a specific function
python3 -c "
import cProfile
from my_module import expensive_function

cProfile.run('expensive_function()', sort='cumulative')
"

line_profiler (line-by-line)

# Install
pip install line_profiler

# Add @profile decorator to functions of interest, then:
kernprof -l -v my_script.py

# Programmatic usage
from line_profiler import LineProfiler

def process_data(data):
    result = []
    for item in data:           # Is this loop the bottleneck?
        transformed = transform(item)
        if validate(transformed):
            result.append(transformed)
    return result

profiler = LineProfiler()
profiler.add_function(process_data)
profiler.enable()
process_data(large_dataset)
profiler.disable()
profiler.print_stats()

Memory profiling (Python)

# memory_profiler
pip install memory_profiler

# Profile memory line-by-line
python3 -m memory_profiler my_script.py

from memory_profiler import profile

@profile
def load_data():
    data = []
    for i in range(1000000):
        data.append({'id': i, 'value': f'item_{i}'})
    return data

# Track memory over time
import tracemalloc

tracemalloc.start()

# ... run code ...

snapshot = tracemalloc.take_snapshot()
top_stats = snapshot.statistics('lineno')
for stat in top_stats[:10]:
    print(stat)

Python benchmarking

import timeit

# Time a statement
result = timeit.timeit('sorted(range(1000))', number=10000)
print(f"sorted: {result:.4f}s for 10000 iterations")

# Compare two approaches
setup = "data = list(range(10000))"
t1 = timeit.timeit('list(filter(lambda x: x % 2 == 0, data))', setup=setup, number=1000)
t2 = timeit.timeit('[x for x in data if x % 2 == 0]', setup=setup, number=1000)
print(f"filter: {t1:.4f}s  |  listcomp: {t2:.4f}s  |  speedup: {t1/t2:.2f}x")

# pytest-benchmark
# pip install pytest-benchmark
# def test_sort(benchmark):
#     benchmark(sorted, list(range(1000)))

Go Profiling

Built-in pprof

// Add to main.go for HTTP-accessible profiling
import (
    "net/http"
    _ "net/http/pprof"
)

func main() {
    go func() {
        http.ListenAndServe("localhost:6060", nil)
    }()
    // ... rest of app
}

# CPU profile (30 seconds)
go tool pprof http://localhost:6060/debug/pprof/profile?seconds=30

# Memory profile
go tool pprof http://localhost:6060/debug/pprof/heap

# Goroutine profile
go tool pprof http://localhost:6060/debug/pprof/goroutine

# Inside pprof interactive mode:
# top 20          - top functions by CPU/memory
# list funcName   - source code with annotations
# web             - open flame graph in browser
# png > out.png   - save call graph as image

Go benchmarks

// math_test.go
func BenchmarkAdd(b *testing.B) {
    for i := 0; i < b.N; i++ {
        Add(42, 58)
    }
}

func BenchmarkSort1000(b *testing.B) {
    data := make([]int, 1000)
    for i := range data {
        data[i] = rand.Intn(1000)
    }
    b.ResetTimer()
    for i := 0; i < b.N; i++ {
        sort.Ints(append([]int{}, data...))
    }
}

# Run benchmarks
go test -bench=. -benchmem ./...

# Compare before/after
go test -bench=. -count=5 ./... > old.txt
# ... make changes ...
go test -bench=. -count=5 ./... > new.txt
go install golang.org/x/perf/cmd/benchstat@latest
benchstat old.txt new.txt

Flame Graphs

Generate flame graphs

# Node.js: 0x (easiest)
npx 0x app.js
# Opens interactive flame graph in browser

# Node.js: clinic.js (comprehensive)
npx clinic flame -- node app.js
npx clinic doctor -- node app.js
npx clinic bubbleprof -- node app.js

# Python: py-spy (sampling profiler, no code changes needed)
pip install py-spy
py-spy record -o flame.svg -- python3 my_script.py

# Profile running Python process
py-spy record -o flame.svg --pid 12345

# Go: built-in
go tool pprof -http=:8080 http://localhost:6060/debug/pprof/profile?seconds=30
# Navigate to "Flame Graph" view

# Linux (any process): perf + flamegraph
perf record -g -p PID -- sleep 30
perf script | stackcollapse-perf.pl | flamegraph.pl > flame.svg

Reading flame graphs

Key concepts:
- X-axis: NOT time. It's alphabetical sort of stack frames. Width = % of samples.
- Y-axis: Stack depth. Top = leaf function (where CPU time is spent).
- Wide bars at the top = hot functions (optimize these first).
- Narrow tall stacks = deep call chains (may indicate excessive abstraction).

What to look for:
1. Wide plateaus at the top → function that dominates CPU time
2. Multiple paths converging to one function → shared bottleneck
3. GC/runtime frames taking significant width → memory pressure
4. Unexpected functions appearing wide → performance bug

Load Testing

curl-based quick test

# Single request timing
curl -o /dev/null -s -w "HTTP %{http_code} | Total: %{time_total}s | TTFB: %{time_starttransfer}s | Connect: %{time_connect}s\
" https://api.example.com/endpoint

# Multiple requests in sequence
for i in $(seq 1 20); do
  curl -o /dev/null -s -w "%{time_total}\
" https://api.example.com/endpoint
done | awk '{sum+=$1; count++; if($1>max)max=$1} END {printf "avg=%.3fs max=%.3fs n=%d\
", sum/count, max, count}'

Apache Bench (ab)

# 100 requests, 10 concurrent
ab -n 100 -c 10 http://localhost:3000/api/endpoint

# With POST data
ab -n 100 -c 10 -p data.json -T application/json http://localhost:3000/api/endpoint

# Key metrics to watch:
# - Requests per second (throughput)
# - Time per request (latency)
# - Percentage of requests served within a certain time (p50, p90, p99)

wrk (modern load testing)

# Install: https://github.com/wg/wrk
# 10 seconds, 4 threads, 100 connections
wrk -t4 -c100 -d10s http://localhost:3000/api/endpoint

# With Lua script for custom requests
wrk -t4 -c100 -d10s -s post.lua http://localhost:3000/api/endpoint

-- post.lua
wrk.method = "POST"
wrk.body   = '{"key": "value"}'
wrk.headers["Content-Type"] = "application/json"

-- Custom request generation
request = function()
  local id = math.random(1, 10000)
  local path = "/api/users/" .. id
  return wrk.format("GET", path)
end

Autocannon (Node.js load testing)

npx autocannon -c 100 -d 10 http://localhost:3000/api/endpoint
npx autocannon -c 100 -d 10 -m POST -b '{"key":"value"}' -H 'Content-Type=application/json' http://localhost:3000/api/endpoint

Database Query Performance

EXPLAIN analysis

# PostgreSQL
psql -c "EXPLAIN (ANALYZE, BUFFERS, FORMAT TEXT) SELECT * FROM orders WHERE user_id = 123;"

# MySQL
mysql -e "EXPLAIN SELECT * FROM orders WHERE user_id = 123;" mydb

# SQLite
sqlite3 mydb.sqlite "EXPLAIN QUERY PLAN SELECT * FROM orders WHERE user_id = 123;"

Slow query detection

# PostgreSQL: enable slow query logging
# In postgresql.conf:
# log_min_duration_statement = 100  (ms)

# MySQL: slow query log
# In my.cnf:
# slow_query_log = 1
# long_query_time = 0.1

# Find queries missing indexes (PostgreSQL)
psql -c "
SELECT schemaname, relname, seq_scan, seq_tup_read,
       idx_scan, idx_tup_fetch,
       seq_tup_read / GREATEST(seq_scan, 1) AS avg_rows_per_scan
FROM pg_stat_user_tables
WHERE seq_scan > 100 AND seq_tup_read / GREATEST(seq_scan, 1) > 1000
ORDER BY seq_tup_read DESC
LIMIT 10;
"

Memory Leak Detection Patterns

Node.js

// Track object counts over time
const v8 = require('v8');

function checkMemory() {
  const heap = v8.getHeapStatistics();
  const usage = process.memoryUsage();
  return {
    heapUsedMB: (usage.heapUsed / 1024 / 1024).toFixed(1),
    heapTotalMB: (usage.heapTotal / 1024 / 1024).toFixed(1),
    rssMB: (usage.rss / 1024 / 1024).toFixed(1),
    externalMB: (usage.external / 1024 / 1024).toFixed(1),
    arrayBuffersMB: (usage.arrayBuffers / 1024 / 1024).toFixed(1),
  };
}

// Sample every 10s, alert on growth
let baseline = process.memoryUsage().heapUsed;
setInterval(() => {
  const current = process.memoryUsage().heapUsed;
  const growthMB = (current - baseline) / 1024 / 1024;
  if (growthMB > 50) {
    console.warn(`Memory grew ${growthMB.toFixed(1)}MB since start`);
    console.warn(checkMemory());
  }
}, 10000);

Common leak patterns

Node.js:
- Event listeners not removed (emitter.on without emitter.off)
- Closures capturing large objects in long-lived scopes
- Global caches without eviction (Map/Set that only grows)
- Unresolved promises accumulating

Python:
- Circular references (use weakref for caches)
- Global lists/dicts that grow unbounded
- File handles not closed (use context managers)
- C extension objects not properly freed

Go:
- Goroutine leaks (goroutine started, never returns)
- Forgotten channel listeners
- Unclosed HTTP response bodies
- Global maps that grow forever

Performance Comparison Script

#!/bin/bash
# perf-compare.sh - Compare performance before/after a change
# Usage: perf-compare.sh <command> [runs]
CMD="${1:?Usage: perf-compare.sh <command> [runs]}"
RUNS="${2:-10}"

echo "Benchmarking: $CMD"
echo "Runs: $RUNS"
echo ""

times=()
for i in $(seq 1 "$RUNS"); do
  start=$(date +%s%N)
  eval "$CMD" > /dev/null 2>&1
  end=$(date +%s%N)
  elapsed=$(echo "scale=3; ($end - $start) / 1000000" | bc)
  times+=("$elapsed")
  printf "  Run %2d: %sms\
" "$i" "$elapsed"
done

echo ""
printf '%s\
' "${times[@]}" | awk '{
  sum += $1
  sumsq += $1 * $1
  if (NR == 1 || $1 < min) min = $1
  if (NR == 1 || $1 > max) max = $1
  count++
} END {
  avg = sum / count
  stddev = sqrt(sumsq/count - avg*avg)
  printf "Results: avg=%.1fms min=%.1fms max=%.1fms stddev=%.1fms (n=%d)\
", avg, min, max, stddev, count
}'

Tips

• Profile before optimizing. Guessing where bottlenecks are is wrong more often than right. Measure first.
• Optimize the hot path. Flame graphs show you exactly which functions consume the most time. A 10% improvement in a function that takes 80% of CPU time is worth more than a 50% improvement in one that takes 2%.
• Memory and CPU are different problems. A memory leak can exist in fast code. A CPU bottleneck can exist in code with stable memory. Profile both independently.
• Benchmark under realistic conditions. Microbenchmarks (empty loops, single-function timing) can be misleading due to JIT optimization, caching, and branch prediction. Use realistic data and workloads.
• p99 matters more than average. An API with 50ms average but 2s p99 has a tail latency problem. Always look at percentiles, not just averages.
• Load test before shipping. ab, wrk, or autocannon for 60 seconds at expected peak traffic reveals problems that unit tests never will.
• GC pauses are real. In Node.js, Python, Go, and Java, garbage collection can cause latency spikes. If flame graphs show significant GC time, reduce allocation pressure (reuse objects, use object pools, avoid unnecessary copies).
• Database queries are usually the bottleneck. Before optimizing application code, run EXPLAIN on your slowest queries. An index can turn a 2-second query into 2ms.