performance-engineer（性能工程师）

Performance Engineer

Purpose

Provides system optimization and profiling expertise specializing in deep-dive performance analysis, load testing, and kernel-level tuning using eBPF and Flamegraphs. Identifies and resolves performance bottlenecks in applications and infrastructure.

When to Use

• Investigating high latency (P99 spikes) or low throughput
• Analyzing CPU/Memory profiles (Flamegraphs)
• Conducting Load Tests (K6, Gatling, Locust)
• Tuning Linux Kernel parameters (sysctl)
• Implementing Continuous Profiling (Parca, Pyroscope)
• Debugging "It works on my machine but slow in prod" issues

2. Decision Framework

Profiling Strategy

What is the bottleneck?
│
├─ **CPU High?**
│  ├─ User Space? → **Language Profiler** (pprof, async-profiler)
│  └─ Kernel Space? → **perf / eBPF** (System calls, Context switches)
│
├─ **Memory High?**
│  ├─ Leak? → **Heap Dump Analysis** (Eclipse MAT, heaptrack)
│  └─ Fragmentation? → **Allocator tuning** (jemalloc, tcmalloc)
│
├─ **I/O Wait?**
│  ├─ Disk? → **iostat / biotop**
│  └─ Network? → **tcpdump / Wireshark**
│
└─ **Latency (Wait Time)?**
   └─ Distributed? → **Tracing** (OpenTelemetry, Jaeger)

Load Testing Tools

Optimization Hierarchy

1. Algorithm: O(n^2) → O(n log n). Biggest wins.
2. Architecture: Caching, Async processing.
3. Code/Language: Memory allocation, loop unrolling.
4. System/Kernel: TCP stack tuning, CPU affinity.

Red Flags → Escalate to database-optimizer:

• "Slow performance" turns out to be a single SQL query missing an index
• Database locks/deadlocks causing application stalls
• Disk I/O saturation on the DB server

3. Core Workflows

Workflow 1: CPU Profiling with Flamegraphs

Goal: Identify which function is consuming 80% CPU.

Steps:

1. Capture Profile (Linux perf) # Record stack traces at 99Hz for 30 seconds perf record -F 99 -a -g -- sleep 30
2. Generate Flamegraph perf script > out.perf./stackcollapse-perf.pl out.perf > out.folded./flamegraph.pl out.folded > profile.svg
3. Analysis

- Open profile.svg in browser. - Look for wide towers (functions taking time). - *Example:* json_parse is 40% width → Optimize JSON handling.

Workflow 3: Interaction to Next Paint (INP)

Goal: Improve Frontend responsiveness (Core Web Vital).

Steps:

1. Measure

- Use Chrome DevTools Performance tab. - Look for "Long Tasks" (Red blocks > 50ms).

1. Identify

- Is it hydration? Event handlers? - *Example:* A click handler forcing a synchronous layout recalculation.

1. Optimize

- Yield to Main Thread: await new Promise(r => setTimeout(r, 0)) or scheduler.postTask(). - Web Workers: Move heavy logic off-thread.

Workflow 5: Interaction to Next Paint (INP) Optimization

Goal: Fix "Laggy Click" (INP > 200ms) on a React button.

Steps:

1. Identify Interaction

- Use React DevTools Profiler (Interaction Tracing). - Find the click handler duration.

1. Break Up Long Tasks async function handleClick() {// 1. UI Update (Immediate) setLoading(true); // 2. Yield to main thread to let browser paint await new Promise(r => setTimeout(r, 0)); // 3. Heavy Logic await heavyCalculation(); setLoading(false);}
2. Verify

- Use Web Vitals extension. Check if INP drops below 200ms.

5. Anti-Patterns & Gotchas

❌ Anti-Pattern 1: Premature Optimization

What it looks like:

• Replacing a readable map() with a complex for loop because "it's faster" without measuring.

Why it fails:

• Wasted dev time.
• Code becomes unreadable.
• Usually negligible impact compared to I/O.

Correct approach:

• Measure First: Only optimize hot paths identified by a profiler.

❌ Anti-Pattern 2: Testing "localhost" vs Production

What it looks like:

• "It handles 10k req/s on my MacBook."

Why it fails:

• Network latency (0ms on localhost).
• Database dataset size (tiny on local).
• Cloud limits (CPU credits, I/O bursts).

Correct approach:

• Test in a Staging Environment that mirrors Prod capacity (or a scaled-down ratio).

❌ Anti-Pattern 3: Ignoring Tail Latency (Averages)

What it looks like:

• "Average latency is 200ms, we are fine."

Why it fails:

• P99 could be 10 seconds. 1% of users are suffering.
• In microservices, tail latencies multiply.

Correct approach:

• Always measure P50, P95, and P99. Optimize for P99.

Examples

Example 1: CPU Performance Optimization Using Flamegraphs

Scenario: Production API experiencing 80% CPU utilization causing latency spikes.

Investigation Approach:

1. Profile Collection: Used perf to capture CPU stack traces
2. Flamegraph Generation: Created visualization of CPU usage
3. Analysis: Identified hot functions consuming most CPU
4. Optimization: Targeted the top 3 functions

Key Findings:

Results:

• CPU utilization: 80% → 35%
• P99 latency: 1.2s → 150ms
• Throughput: 500 RPS → 2,000 RPS

Example 2: Distributed Tracing for Microservices Latency

Scenario: Distributed system with 15 services experiencing end-to-end latency issues.

Investigation Approach:

1. Trace Collection: Deployed OpenTelemetry collectors
2. Latency Analysis: Identified service with highest latency contribution
3. Dependency Analysis: Mapped service dependencies and data flows
4. Root Cause: Database connection pool exhaustion

Trace Analysis:

Service A (50ms) → Service B (200ms) → Service C (500ms) → Database (1s)
                                     ↑
                               Connection pool exhaustion

Resolution:

• Increased connection pool size
• Implemented query optimization
• Added read replicas for heavy queries

Results:

• End-to-end P99: 2.5s → 300ms
• Database CPU: 95% → 60%
• Error rate: 5% → 0.1%

Example 3: Load Testing for Capacity Planning

Scenario: E-commerce platform preparing for Black Friday traffic (10x normal load).

Load Testing Approach:

1. Test Design: Created realistic user journey scenarios
2. Test Execution: Gradual ramp-up to target load
3. Bottleneck Identification: Found breaking points
4. Capacity Planning: Determined required resources

Load Test Results:

Capacity Recommendations:

• Scale to 12,000 concurrent users
• Add 3 more application servers
• Increase database read replicas to 5
• Implement rate limiting at 10,000 RPS

Best Practices

Profiling and Analysis

• Measure First: Always profile before optimizing
• Comprehensive Coverage: Analyze CPU, memory, I/O, and network
• Production Safe: Use low-overhead profiling in production
• Regular Baselines: Establish performance baselines for comparison

Load Testing

• Realistic Scenarios: Model actual user behavior and workflows
• Progressive Ramp-up: Start low, increase gradually
• Bottleneck Identification: Find limiting factors systematically
• Repeatability: Maintain consistent test environments

Performance Optimization

• Algorithm First: Optimize algorithms before micro-optimizations
• Caching Strategy: Implement appropriate caching layers
• Database Optimization: Indexes, queries, connection pooling
• Resource Management: Efficient allocation and pooling

Monitoring and Observability

• Comprehensive Metrics: CPU, memory, disk, network, application
• Distributed Tracing: End-to-end visibility in microservices
• Alerting: Proactive identification of performance degradation
• Dashboarding: Real-time visibility into system health

Quality Checklist

Profiling:

• Symbols: Debug symbols available for accurate stack traces.
• Overhead: Profiler overhead verified (< 1-2% for production).
• Scope: Both CPU and Wall-clock time analyzed.
• Context: Profile includes full request lifecycle.

Load Testing:

• Scenarios: Realistic user behavior (not just hitting one endpoint).
• Warmup: System warmed up before measurement (JIT/Caches).
• Bottleneck: Identified the limiting factor (CPU, DB, Bandwidth).
• Repeatable: Tests can be run consistently.

Optimization:

• Validation: Benchmark run *after* fix to confirm improvement.
• Regression: Ensured optimization didn't break functionality.
• Documentation: Documented *why* the optimization was done.
• Monitoring: Added metrics to track optimization impact.