2nd/10_Wiki/Topic_Programming/DevOps_and_Security/Efficiency.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-efficiency	Efficiency	10_Wiki/Topics	verified	self	Performance Efficiency Resource Efficiency Cost Efficiency	none	A	0.88	applied	performance efficiency optimization sre cost		2026-05-10	pending	language framework python prometheus

Efficiency

매 한 줄

"매 useful output / resource consumed — 매 latency, throughput, $cost, watt 매 dimension 별 측정". 매 1972 Knuth "premature optimization" warning 후 신중함이 default, 매 2026 cloud cost + carbon footprint + energy efficiency 가 first-class metric.

매 핵심

매 Efficiency dimensions

• Time: latency p50/p95/p99, throughput RPS.
• Space: memory RSS, disk IOPS, network bytes.
• Money: $/request, $/MAU.
• Energy: watt/op, gCO2eq/request.

매 측정 → 개선 cycle

1. Profile: hotspot 의 identify (flamegraph).
2. Hypothesize: bottleneck 의 type (CPU? IO? Lock?).
3. Optimize: targeted change.
4. Verify: A/B with baseline, metric 의 statistical sig.

매 응용

1. API: cold-start 의 reduce.
2. ML inference: quantization, batching, KV cache.
3. CI: cache hit rate 의 maximize.

💻 패턴

Latency histogram (Prometheus)

from prometheus_client import Histogram
LATENCY = Histogram('http_request_duration_seconds', 'HTTP latency',
                    buckets=[0.005, 0.01, 0.025, 0.05, 0.1, 0.25, 0.5, 1, 2.5, 5])

@LATENCY.time()
def handle(req): ...

Cost-per-request rollup

-- BigQuery
SELECT
  service,
  SUM(billable_seconds * cpu_cost_per_sec) AS compute_usd,
  COUNT(*) AS requests,
  SUM(billable_seconds * cpu_cost_per_sec) / COUNT(*) AS usd_per_req
FROM service_metrics
WHERE _PARTITIONDATE = CURRENT_DATE() - 1
GROUP BY service
ORDER BY usd_per_req DESC;

CPU flamegraph (py-spy)

# 매 production-safe sampling profiler
py-spy record -o flame.svg -d 60 -p $(pgrep -f gunicorn)
py-spy top -p $(pgrep -f gunicorn)

Memory profiling (memray)

memray run --live ./app.py
memray flamegraph output.bin -o memflame.html

Async IO efficiency

import asyncio, httpx

# BAD — sequential
async def fetch_seq(urls):
    async with httpx.AsyncClient() as c:
        return [await c.get(u) for u in urls]

# GOOD — concurrent
async def fetch_par(urls):
    async with httpx.AsyncClient() as c:
        return await asyncio.gather(*[c.get(u) for u in urls])

Carbon-aware scheduling

import httpx
async def carbon_intensity(region: str) -> float:
    r = await httpx.AsyncClient().get(
        f"https://api.electricitymaps.com/v3/carbon-intensity/latest?zone={region}",
        headers={"auth-token": "TOKEN"})
    return r.json()["carbonIntensity"]  # gCO2eq/kWh

# 매 batch job 매 low-carbon window 의 schedule
async def maybe_run(region):
    ci = await carbon_intensity(region)
    if ci < 200: await run_batch()
    else: await asyncio.sleep(900)

LLM inference batching

from vllm import LLM, SamplingParams
llm = LLM(model="meta-llama/Llama-3.3-70B-Instruct",
          tensor_parallel_size=4,
          max_num_seqs=256)  # 매 batch 의 throughput
sp = SamplingParams(max_tokens=256)
outs = llm.generate(prompts, sp)  # 매 single forward pass 로 batch

Cache efficiency dashboard

# Prometheus query — cache hit ratio
sum(rate(cache_hits_total[5m])) /
(sum(rate(cache_hits_total[5m])) + sum(rate(cache_misses_total[5m])))

매 결정 기준

상황	Approach
Latency-critical (real-time)	tail-latency optimize, drop p99 outliers
Throughput (batch)	parallelism, vectorize
Cost 제약	spot instance, autoscale, cache
Green ops	carbon-aware scheduling, region selection

기본값: 매 profile-first, 매 measure both before/after, 매 single-dimension fixation 매 X.

🔗 Graph

• 부모: Site Reliability Engineering
• 변형: Latency
• 응용: Flame_Graphs · Profiling
• Adjacent: FinOps · Green Software

🤖 LLM 활용

언제: profile output → bottleneck classification, optimization 후보 ranking. 언제 X: 매 micro-benchmark 매 LLM 의 single-shot 평가 X — 매 측정 우선.

❌ 안티패턴

• Premature optimization: profile 없이 optimize.
• Single-metric obsession: latency 만 보고 cost 폭증.
• Synthetic benchmark: production traffic shape 무시.
• No baseline: 매 before 측정 없이 "fast" 주장.

🧪 검증 / 중복

• Verified (SRE Workbook ch.4, AWS Well-Architected Performance pillar).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — efficiency 4 dimension + measurement loop