File size: 5,706 Bytes
23a7a20 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 | #!/usr/bin/env python3
"""Compute ANNS workload statistics for evaluation."""
import os
import json
import pandas as pd
import numpy as np
from transformers import AutoTokenizer
import argparse
def parse_pipeline_pool(pool_str: str):
"""Parse pipeline pool string to extract document IDs."""
pool_str = pool_str.strip('()')
if not pool_str:
return []
return [doc_id.strip() for doc_id in pool_str.split(',')]
def main():
parser = argparse.ArgumentParser(description="Compute ANNS workload statistics")
parser.add_argument("--corpus-prefix",
default="retrieved_corpus_content",
help="Prefix for corpus content part files")
parser.add_argument("--query-map",
default="query_trace_map_5k.json",
help="Path to query trace map JSON file")
parser.add_argument("--trace-dir",
default="res",
help="Directory containing trace CSV files")
parser.add_argument("--max-queries",
type=int,
default=500,
help="Maximum number of queries to process")
parser.add_argument("--tokenizer-model",
default="meta-llama/Llama-3.1-8B-Instruct",
help="HuggingFace tokenizer model")
parser.add_argument("--output-dir",
default="tables",
help="Output directory for statistics file")
args = parser.parse_args()
# Load corpus
print("Loading corpus content...")
corpus_content = {}
part_num = 0
while True:
part_file = f"{args.corpus_prefix}.{part_num}.json"
if not os.path.exists(part_file):
break
print(f" Loading {part_file}...")
with open(part_file, 'r') as f:
part_data = json.load(f)
corpus_content.update(part_data)
part_num += 1
print(f"Loaded {len(corpus_content)} documents")
# Load query map
with open(args.query_map, 'r') as f:
query_trace_map = json.load(f)
# Load tokenizer
print("Loading tokenizer...")
try:
tokenizer = AutoTokenizer.from_pretrained(
args.tokenizer_model, local_files_only=True)
except:
tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_model)
# Process queries
query_items = list(query_trace_map.items())[:args.max_queries]
print(f"Processing {len(query_items)} queries...")
total_query_tokens = []
query_durations = []
for query_id, query_info in query_items:
# Read trace file
trace_path = os.path.join(args.trace_dir, query_info['trace_file'])
if not os.path.exists(trace_path):
continue
try:
df = pd.read_csv(trace_path)
if df.empty:
continue
# Get duration
start_time_us = df['StartTime_us'].iloc[0]
end_time_us = df['EndTime_us'].iloc[-1]
duration_secs = (end_time_us - start_time_us) / 1e6
query_durations.append(duration_secs)
# Get pipeline pool and tokenize
final_row = df.iloc[-1]
pipeline_pool_str = str(final_row['PipelinePool']).strip('()')
if pipeline_pool_str:
doc_ids = [d.strip() for d in pipeline_pool_str.split(',')]
else:
doc_ids = []
# Tokenize query
query_tokens = len(
tokenizer.encode(query_info['query'],
truncation=False,
add_special_tokens=True))
# Tokenize documents
total_doc_tokens = 0
for doc_id in doc_ids:
if doc_id not in corpus_content:
continue
doc_text = corpus_content[doc_id]
doc_tokens = len(
tokenizer.encode(doc_text,
truncation=False,
add_special_tokens=True))
total_doc_tokens += doc_tokens
total_tokens = query_tokens + total_doc_tokens
total_query_tokens.append(total_tokens)
except Exception as e:
continue
# Compute statistics and save to file
os.makedirs(args.output_dir, exist_ok=True)
output_file = os.path.join(args.output_dir, "workload_stats_anns.txt")
with open(output_file, 'w') as f:
f.write("\n" + "=" * 70 + "\n")
f.write("ANNS WORKLOAD STATISTICS\n")
f.write("=" * 70 + "\n")
if total_query_tokens:
total_query_tokens = np.array(total_query_tokens)
f.write(f"\nTotal Tokens per Query (n={len(total_query_tokens)})\n")
f.write(f" Mean: {total_query_tokens.mean():.0f} tokens\n")
f.write(f" P50: {np.percentile(total_query_tokens, 50):.0f} tokens\n")
f.write(f" P75: {np.percentile(total_query_tokens, 75):.0f} tokens\n")
f.write(f" P95: {np.percentile(total_query_tokens, 95):.0f} tokens\n")
if query_durations:
query_durations = np.array(query_durations)
f.write(f"\nQuery Duration (n={len(query_durations)})\n")
f.write(f" Mean: {query_durations.mean():.3f} seconds\n")
f.write(f" P50: {np.percentile(query_durations, 50):.3f} seconds\n")
f.write(f" P75: {np.percentile(query_durations, 75):.3f} seconds\n")
f.write(f" P95: {np.percentile(query_durations, 95):.3f} seconds\n")
f.write("=" * 70 + "\n")
if __name__ == "__main__":
main()
|