Skip to content
Back to skills

M10 Performance

ASecurity

CRITICAL: Use for performance optimization. Triggers: performance, optimization, benchmark, profiling, flamegraph, criterion, slow, fast, allocation, cache, SIMD, make it faster, 性能优化, 基准测试

  • 18 stars
  • 0 votes
  • 0 copies
  • 4 views
  • Added August 31, 2026
businessgoperformance

Security analysis

A100/100

Pro scans all 2 files and shows the line behind each finding

Scanned August 31, 2026

npx -y skills add k1-c/linear-tui --skill m10-performance --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of M10 Performance?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for M10 Performance
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/k1-c-m10-performance/badge)](https://www.skillsdirectory.com/skills/k1-c-m10-performance)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: m10-performance
description: "CRITICAL: Use for performance optimization. Triggers: performance, optimization, benchmark, profiling, flamegraph, criterion, slow, fast, allocation, cache, SIMD, make it faster, 性能优化, 基准测试"
user-invocable: false
---

# Performance Optimization

> **Layer 2: Design Choices**

## Core Question

**What's the bottleneck, and is optimization worth it?**

Before optimizing:
- Have you measured? (Don't guess)
- What's the acceptable performance?
- Will optimization add complexity?

---

## Performance Decision → Implementation

| Goal | Design Choice | Implementation |
|------|---------------|----------------|
| Reduce allocations | Pre-allocate, reuse | `with_capacity`, object pools |
| Improve cache | Contiguous data | `Vec`, `SmallVec` |
| Parallelize | Data parallelism | `rayon`, threads |
| Avoid copies | Zero-copy | References, `Cow<T>` |
| Reduce indirection | Inline data | `smallvec`, arrays |

---

## Thinking Prompt

Before optimizing:

1. **Have you measured?**
   - Profile first → flamegraph, perf
   - Benchmark → criterion, cargo bench
   - Identify actual hotspots

2. **What's the priority?**
   - Algorithm (10x-1000x improvement)
   - Data structure (2x-10x)
   - Allocation (2x-5x)
   - Cache (1.5x-3x)

3. **What's the trade-off?**
   - Complexity vs speed
   - Memory vs CPU
   - Latency vs throughput

---

## Trace Up ↑

To domain constraints (Layer 3):

```
"How fast does this need to be?"
    ↑ Ask: What's the performance SLA?
    ↑ Check: domain-* (latency requirements)
    ↑ Check: Business requirements (acceptable response time)
```

| Question | Trace To | Ask |
|----------|----------|-----|
| Latency requirements | domain-* | What's acceptable response time? |
| Throughput needs | domain-* | How many requests per second? |
| Memory constraints | domain-* | What's the memory budget? |

---

## Trace Down ↓

To implementation (Layer 1):

```
"Need to reduce allocations"
    ↓ m01-ownership: Use references, avoid clone
    ↓ m02-resource: Pre-allocate with_capacity

"Need to parallelize"
    ↓ m07-concurrency: Choose rayon or threads
    ↓ m07-concurrency: Consider async for I/O-bound

"Need cache efficiency"
    ↓ Data layout: Prefer Vec over HashMap when possible
    ↓ Access patterns: Sequential over random access
```

---

## Quick Reference

| Tool | Purpose |
|------|---------|
| `cargo bench` | Micro-benchmarks |
| `criterion` | Statistical benchmarks |
| `perf` / `flamegraph` | CPU profiling |
| `heaptrack` | Allocation tracking |
| `valgrind` / `cachegrind` | Cache analysis |

## Optimization Priority

```
1. Algorithm choice     (10x - 1000x)
2. Data structure       (2x - 10x)
3. Allocation reduction (2x - 5x)
4. Cache optimization   (1.5x - 3x)
5. SIMD/Parallelism     (2x - 8x)
```

## Common Techniques

| Technique | When | How |
|-----------|------|-----|
| Pre-allocation | Known size | `Vec::with_capacity(n)` |
| Avoid cloning | Hot paths | Use references or `Cow<T>` |
| Batch operations | Many small ops | Collect then process |
| SmallVec | Usually small | `smallvec::SmallVec<[T; N]>` |
| Inline buffers | Fixed-size data | Arrays over Vec |

---

## Common Mistakes

| Mistake | Why Wrong | Better |
|---------|-----------|--------|
| Optimize without profiling | Wrong target | Profile first |
| Benchmark in debug mode | Meaningless | Always `--release` |
| Use LinkedList | Cache unfriendly | `Vec` or `VecDeque` |
| Hidden `.clone()` | Unnecessary allocs | Use references |
| Premature optimization | Wasted effort | Make it work first |

---

## Anti-Patterns

| Anti-Pattern | Why Bad | Better |
|--------------|---------|--------|
| Clone to avoid lifetimes | Performance cost | Proper ownership |
| Box everything | Indirection cost | Stack when possible |
| HashMap for small sets | Overhead | Vec with linear search |
| String concat in loop | O(n^2) | `String::with_capacity` or `format!` |

---

## Related Skills

| When | See |
|------|-----|
| Reducing clones | m01-ownership |
| Concurrency options | m07-concurrency |
| Smart pointer choice | m02-resource |
| Domain requirements | domain-* |

Files in this skill

  • SKILL.md4.1 KB
  • patterns/optimization-guide.md6.9 KB

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.