CPU 임베딩 벤치마크 도구 — 8종 CPU 성능 비교 자동화

# cpu_embedding_bench.py

import os
import time
import json
import platform
from sentence_transformers import SentenceTransformer

MODELS = [
    "all-MiniLM-L6-v2",
    "BAAI/bge-base-en-v1.5",
    "intfloat/multilingual-e5-large",
]

# 테스트 데이터 (1,000개 문장)
sentences = [
    f"This is test sentence number {i} for "
    f"embedding benchmark."
    for i in range(1000)
]

def benchmark_model(model_name, batch_size=32, runs=3):
    model = SentenceTransformer(model_name)

    # 웜업 (첫 실행은 느림)
    model.encode(sentences[:10], batch_size=batch_size)

    times = []
    for _ in range(runs):
        start = time.perf_counter()
        model.encode(sentences, batch_size=batch_size)
        elapsed = time.perf_counter() - start
        times.append(elapsed)

    median_time = sorted(times)[len(times) // 2]
    throughput = len(sentences) / median_time

    return {
        "model": model_name,
        "median_seconds": round(median_time, 2),
        "sentences_per_sec": round(throughput, 1),
    }

results = {
    "cpu": platform.processor(),
    "cores": os.cpu_count(),
    "benchmarks": [
        benchmark_model(m) for m in MODELS
    ],
}

print(json.dumps(results, indent=2))

# 실행 방법

# 의존성 설치
pip install sentence-transformers torch

# 벤치마크 실행 (결과를 JSON으로 저장)
python cpu_embedding_bench.py > results.json

# 또는 특정 CPU 코어 수로 제한
OMP_NUM_THREADS=4 python cpu_embedding_bench.py

# remote-bench.sh - 여러 서버에서 벤치마크 원격 실행

#!/bin/bash

SERVERS=(
  "server-a"
  "server-b"
  "server-c"
  "minipc-a"
)

SCRIPT_PATH="/opt/bench/cpu_embedding_bench.py"
OUTPUT_DIR="./results"

mkdir -p "$OUTPUT_DIR"

for server in "${SERVERS[@]}"; do
  echo "[$server] 벤치마크 시작..."

  ssh "$server" "python3 $SCRIPT_PATH" \
    > "$OUTPUT_DIR/$server.json" 2>/dev/null

  if [ $? -eq 0 ]; then
    echo "[$server] 완료"
  else
    echo "[$server] 실패"
  fi
done

echo "모든 서버 벤치마크 완료"
echo "결과: $OUTPUT_DIR/"

# 단일 서버 결과 예시 (server-a.json)

{
  "cpu": "AMD Ryzen 9 9950X3D",
  "cores": 32,
  "benchmarks": [
    {
      "model": "all-MiniLM-L6-v2",
      "median_seconds": 1.18,
      "sentences_per_sec": 847.3
    },
    {
      "model": "BAAI/bge-base-en-v1.5",
      "median_seconds": 3.42,
      "sentences_per_sec": 292.4
    },
    {
      "model": "intfloat/multilingual-e5-large",
      "median_seconds": 8.75,
      "sentences_per_sec": 114.3
    }
  ]
}

# merge_results.py - 여러 서버 결과 병합

import json
import glob

results = []
for path in sorted(glob.glob("results/*.json")):
    with open(path) as f:
        data = json.load(f)
        data["server"] = path.split("/")[-1].replace(
            ".json", ""
        )
        results.append(data)

# 모델별로 정렬된 비교표 생성
for model_idx in range(len(results[0]["benchmarks"])):
    model_name = results[0]["benchmarks"][model_idx]["model"]
    print(f"\n=== {model_name} ===")
    ranked = sorted(
        results,
        key=lambda r: r["benchmarks"][model_idx][
            "sentences_per_sec"
        ],
        reverse=True,
    )
    for r in ranked:
        bench = r["benchmarks"][model_idx]
        print(
            f"  {r['cpu']:30s}"
            f"  {bench['sentences_per_sec']:8.1f} sent/s"
        )