2nd/10_Wiki/Topics/Frontend/Cache miss rates.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-cache-miss-rates	Cache Miss Rates	10_Wiki/Topics	verified	self	Cache Miss Rate Cache Hit Rate Cache Efficiency	none	A	0.9	applied	performance cache profiling cpu memory		2026-05-10	pending	language framework agnostic agnostic

Cache Miss Rates

매 한 줄

"매 cache lookup 의 fail (data 의 cache 외 fetch) ratio". CPU L1/L2/L3 cache, browser HTTP cache, application-level memoization, CDN edge cache 의 universal metric — miss / (hit + miss). 2026 perspective: Apple Silicon M4 의 huge L2 (16MB+), 매 Cloudflare/Fastly tiered cache, 매 LLM prompt-cache (Anthropic) 의 cost-driven optimization.

매 핵심

매 hierarchy (CPU)

• L1: 32-128KB / core, ~1-4 cycles.
• L2: 256KB-2MB / core, ~10-15 cycles.
• L3: 8-128MB shared, ~40 cycles.
• DRAM: >100ns, ~200-300 cycles. 매 miss 의 huge cost.

매 miss types (3C)

• Compulsory (cold): 매 first access.
• Capacity: 매 working set > cache size.
• Conflict: 매 set-associative collision.
• (+) Coherence: multi-core invalidation.

매 measurement

• CPU: perf stat -e cache-misses,cache-references (Linux).
• HTTP: cf-cache-status: HIT/MISS headers.
• App: hit/miss counter on cache wrapper.

매 응용

1. 매 hot loop 의 data layout (SoA vs AoS) tuning.
2. 매 CDN cache key strategy.
3. 매 LLM prompt-cache 의 prefix stability.

💻 패턴

Pattern 1 — 매 cache-friendly layout (SoA)

// 매 BAD (AoS) — 매 padding / strided access
const particles = [{ x: 0, y: 0, vx: 1, vy: 1 }, /* ... */];

// 매 GOOD (SoA) — 매 sequential, cache-line dense
const xs = new Float32Array(N), ys = new Float32Array(N);
const vxs = new Float32Array(N), vys = new Float32Array(N);
for (let i = 0; i < N; i++) {
  xs[i] += vxs[i];
  ys[i] += vys[i];
}

Pattern 2 — 매 LRU cache (memoization)

class LRU<K, V> {
  private map = new Map<K, V>();
  constructor(private capacity: number) {}

  get(key: K): V | undefined {
    if (!this.map.has(key)) return;
    const v = this.map.get(key)!;
    this.map.delete(key);
    this.map.set(key, v); // refresh
    return v;
  }
  set(key: K, val: V) {
    if (this.map.has(key)) this.map.delete(key);
    else if (this.map.size >= this.capacity) {
      this.map.delete(this.map.keys().next().value!);
    }
    this.map.set(key, val);
  }
}

Pattern 3 — 매 hit rate metric

class CountingCache<K, V> {
  private hits = 0;
  private misses = 0;
  constructor(private inner: Map<K, V>, private load: (k: K) => V) {}

  get(k: K): V {
    if (this.inner.has(k)) { this.hits++; return this.inner.get(k)!; }
    this.misses++;
    const v = this.load(k);
    this.inner.set(k, v);
    return v;
  }
  hitRate(): number { return this.hits / (this.hits + this.misses || 1); }
}

Pattern 4 — 매 HTTP cache header

GET /api/posts/123
Cache-Control: public, max-age=300, stale-while-revalidate=86400
ETag: "abc123"

# CDN response
cf-cache-status: HIT
age: 42

Pattern 5 — 매 stable key (CDN)

// 매 BAD — query order varies
fetch(`/api/list?b=2&a=1`);
fetch(`/api/list?a=1&b=2`); // 매 different cache key

// 매 GOOD — canonical
const params = new URLSearchParams();
[...Object.entries(args)].sort().forEach(([k, v]) => params.set(k, v));
fetch(`/api/list?${params}`);

Pattern 6 — 매 LLM prompt cache (Anthropic)

# 매 stable system prefix → cache hit (90% cost reduction)
client.messages.create(
    model="claude-opus-4-7",
    system=[
        {"type": "text", "text": LARGE_STABLE_PREFIX,
         "cache_control": {"type": "ephemeral"}},
    ],
    messages=[{"role": "user", "content": query}],
)

Pattern 7 — 매 CPU perf measure

perf stat -e cache-references,cache-misses,L1-dcache-load-misses ./app
# 매 miss rate = misses / references

Pattern 8 — 매 prefetch hint

// 매 link prefetch (browser)
<link rel="prefetch" href="/next-page" as="document">

// 매 software prefetch (WASM/native via SIMD intrinsics)

매 결정 기준

상황	Approach
매 hot loop slow	profile cache-misses, switch to SoA.
매 CDN low hit rate	normalize key, raise max-age.
매 LLM cost high	prompt-cache stable prefix.
매 working set > cache	partition (tiling), reduce.
매 frequent-but-changing	LRU + TTL.

기본값: 매 measure first (perf, CDN headers, app counter), 매 90%+ hit rate target on hot caches.

🔗 Graph

• 부모: Performance
• 변형: CDN · Memoization
• 응용: Prompt Cache
• Adjacent: CPU Overhead · Buffer Allocation · Batching

🤖 LLM 활용

언제: 매 cache strategy design, hit rate target setting, prompt-cache prefix structuring. 언제 X: 매 hardware-specific cache line size assumption — vendor docs 의.

❌ 안티패턴

• 매 cache everything: 매 cold data cache space waste.
• 매 unstable cache key: 매 hit rate near zero.
• 매 no eviction: unbounded memory.
• 매 measuring without baseline: 매 hit rate alone meaningless — 매 latency / cost outcome 의.
• 매 caching mutable data without invalidation: stale read bug.

🧪 검증 / 중복

• Reference (Hennessy & Patterson, Cloudflare cache docs, Anthropic prompt caching docs).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — cache hierarchy + LRU + HTTP + prompt-cache patterns