AINativeBench / data /processed /RQ3 /analyze_performance.py
王子睿
Update processed data
c4e7970 7 months ago
30.7 kB
#!/usr/bin/env python3
"""
Performance Breakdown Analysis Script
This script analyzes execution_path.md files from different models and tasks,
extracting time and token statistics grouped by task, architecture, model, and agent.
"""

import os
import re
import csv
import json
import numpy as np
from pathlib import Path
from collections import defaultdict
from typing import Dict, List, Tuple, Optional
import logging

# Setup logging
logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s - %(levelname)s - %(message)s",
    handlers=[logging.StreamHandler()],
)
logger = logging.getLogger(__name__)


class ExecutionPathParser:
    """Parser for execution_path.md files"""

    def __init__(self, file_path: str):
        self.file_path = file_path
        self.model = None
        self.task = None
        self.architecture = None
        self.total_time = None
        self.total_tokens = {
            "input": 0,
            "output": 0,
            "reasoning": 0,
            "result": 0,
            "total": 0,
        }
        self.agent_stats = defaultdict(
            lambda: {
                "time": [],
                "tokens": {
                    "input": [],
                    "output": [],
                    "reasoning": [],
                    "result": [],
                    "total": [],
                },
            }
        )

    def extract_metadata_from_path(self):
        """Extract model, task, and architecture from file path"""
        try:
            parts = Path(self.file_path).parts
            # Find RESULTS index
            results_idx = parts.index("RESULTS")

            # Model is the next directory after RESULTS
            self.model = parts[results_idx + 1]

            # Task name is the next directory
            task_full = parts[results_idx + 2]

            # Extract architecture from task name
            if "-MCP" in task_full:
                self.architecture = "MCP"
                self.task = task_full.replace("-MCP", "")
            elif "-A2A_mix" in task_full:
                self.architecture = "A2A_mix"
                self.task = task_full.replace("-A2A_mix", "")
            elif "-A2A" in task_full:
                self.architecture = "A2A"
                self.task = task_full.replace("-A2A", "")
            else:
                # Check in file content for architecture info
                self.task = task_full
                self.architecture = "Unknown"

            logger.debug(
                f"Extracted: model={self.model}, task={self.task}, arch={self.architecture}"
            )
            return True
        except Exception as e:
            logger.error(f"Failed to extract metadata from path {self.file_path}: {e}")
            return False

    def parse_tokens_time(self, line: str) -> Tuple[Optional[Dict], Optional[float]]:
        """
        Parse tokens and time from a line like:
        [SPAN] name [∑ tokens: (input→output [REASONING:reasoning, OUTPUT:result], total: total), time: 123.45s]
        """
        tokens_dict = None
        time_val = None

        # Extract tokens
        token_pattern = r"\[∑ tokens: \((\d+)→(\d+) \[REASONING:(\d+), OUTPUT:(\d+)\], total: (\d+)\)"
        token_match = re.search(token_pattern, line)
        if token_match:
            tokens_dict = {
                "input": int(token_match.group(1)),
                "output": int(token_match.group(2)),
                "reasoning": int(token_match.group(3)),
                "result": int(token_match.group(4)),
                "total": int(token_match.group(5)),
            }
        else:
            # Try simpler pattern for LLM calls
            token_pattern2 = (
                r"\((\d+)→(\d+) \[REASONING:(\d+), OUTPUT:(\d+)\], total: (\d+)\)"
            )
            token_match2 = re.search(token_pattern2, line)
            if token_match2:
                tokens_dict = {
                    "input": int(token_match2.group(1)),
                    "output": int(token_match2.group(2)),
                    "reasoning": int(token_match2.group(3)),
                    "result": int(token_match2.group(4)),
                    "total": int(token_match2.group(5)),
                }

        # Extract time
        time_pattern = r"time: ([\d.]+)s\]"
        time_match = re.search(time_pattern, line)
        if time_match:
            time_val = float(time_match.group(1))

        return tokens_dict, time_val

    def extract_agent_name(self, line: str) -> Optional[str]:
        """Extract agent name from a line"""
        # Pattern for agent execution
        agent_pattern = r"\[AGENT\] (.+?)\._execute_core"
        match = re.search(agent_pattern, line)
        if match:
            return match.group(1)

        # Pattern for agent invocation
        agent_pattern2 = r"\[AGENT\] invoke_agent (\w+)"
        match2 = re.search(agent_pattern2, line)
        if match2:
            return match2.group(1)

        # Pattern for agent creation or general agent
        agent_pattern3 = r"\[AGENT\] (?:create_agent )?(\w+)"
        match3 = re.search(agent_pattern3, line)
        if match3 and "tokens:" in line:  # Only if it has token info
            return match3.group(1)

        return None

    def parse_file(self) -> bool:
        """Parse the execution_path.md file"""
        try:
            if not self.extract_metadata_from_path():
                return False

            with open(self.file_path, "r", encoding="utf-8") as f:
                content = f.read()

            # Check if architecture is still Unknown, try to extract from content
            if self.architecture == "Unknown":
                if "A2A_mix" in content:
                    self.architecture = "A2A_mix"
                elif (
                    "Project Type**: A2A" in content or "Project Type**: A2A" in content
                ):
                    self.architecture = "A2A"
                elif "MCP" in content:
                    self.architecture = "MCP"

            # Find the execution tree section
            tree_start = content.find("## Execution Path Tree")
            if tree_start == -1:
                logger.warning(f"No execution tree found in {self.file_path}")
                return False

            tree_section = content[tree_start:]
            lines = tree_section.split("\n")

            # Find the first line with total time (usually the root SPAN)
            # Improvement: support SPAN lines with an error marker prefix (❌)
            for line in lines:
                # Remove the error marker prefix to match correctly
                clean_line = line.replace("❌ ", "")
                if (
                    "[SPAN]" in clean_line or "[Chain]" in clean_line
                ) and "time:" in clean_line:
                    tokens, time = self.parse_tokens_time(clean_line)
                    if time and self.total_time is None:
                        self.total_time = time
                    if tokens:
                        self.total_tokens = tokens
                    break

            # If still not found, try any line that contains time information
            if self.total_time is None:
                for line in lines:
                    clean_line = line.replace("❌ ", "")
                    if "time:" in clean_line and "[∑" in clean_line:
                        tokens, time = self.parse_tokens_time(clean_line)
                        if time:
                            self.total_time = time
                            logger.info(
                                f"Extracted time from error/alternative node: {self.file_path}"
                            )
                            if tokens:
                                self.total_tokens = tokens
                            break

            # Parse all agent lines
            # Improvement: support agent lines with an error marker prefix
            for line in lines:
                if "[AGENT]" in line:
                    # Remove the error marker to parse correctly
                    clean_line = line.replace("❌ ", "")
                    agent_name = self.extract_agent_name(clean_line)
                    tokens, time = self.parse_tokens_time(clean_line)

                    if agent_name:
                        if time is not None:
                            self.agent_stats[agent_name]["time"].append(time)
                        if tokens:
                            for key in [
                                "input",
                                "output",
                                "reasoning",
                                "result",
                                "total",
                            ]:
                                self.agent_stats[agent_name]["tokens"][key].append(
                                    tokens[key]
                                )

            # Improvement: even if total time cannot be extracted, treat it as partially successful if agent stats exist
            if self.total_time is None:
                # Check whether we at least have agent data
                if not self.agent_stats:
                    logger.warning(
                        f"Could not extract total time or agent stats from {self.file_path}"
                    )
                    return False
                else:
                    logger.warning(
                        f"Could not extract total time, but found agent stats in {self.file_path}"
                    )
                    # Use the sum of agent times as an approximate total time
                    all_agent_times = []
                    for agent_data in self.agent_stats.values():
                        all_agent_times.extend(agent_data["time"])
                    if all_agent_times:
                        self.total_time = sum(all_agent_times)
                        logger.info(
                            f"Approximated total time from agent stats: {self.total_time}s"
                        )

            logger.info(
                f"Successfully parsed {self.file_path}: {self.model}/{self.task}/{self.architecture}, time={self.total_time}s"
            )
            return True

        except Exception as e:
            logger.error(f"Failed to parse {self.file_path}: {e}", exc_info=True)
            return False


class PerformanceAnalyzer:
    """Analyzer for performance statistics"""

    def __init__(self, results_dir: str):
        self.results_dir = results_dir
        self.data = []
        self.failed_files = []

    def find_all_execution_paths(self) -> List[str]:
        """Find all execution_path.md files"""
        execution_paths = []
        for root, dirs, files in os.walk(self.results_dir):
            # Skip the RQ- directories
            if "RQ-" in root:
                continue
            if "execution_path.md" in files:
                execution_paths.append(os.path.join(root, "execution_path.md"))

        logger.info(f"Found {len(execution_paths)} execution_path.md files")
        return execution_paths

    def parse_all_files(self):
        """Parse all execution path files"""
        files = self.find_all_execution_paths()

        for file_path in files:
            parser = ExecutionPathParser(file_path)
            if parser.parse_file():
                self.data.append(parser)
            else:
                self.failed_files.append(file_path)

        logger.info(
            f"Successfully parsed {len(self.data)} files, {len(self.failed_files)} failed"
        )

    def calculate_statistics(self, values: List[float]) -> Dict[str, float]:
        """Calculate mean, P90, P99, CV (for time metrics)"""
        if not values:
            return {"mean": 0, "p90": 0, "p99": 0, "cv": 0}

        mean_val = np.mean(values)
        std_val = np.std(values, ddof=1) if len(values) > 1 else 0
        cv = (std_val / mean_val * 100) if mean_val > 0 else 0

        return {
            "mean": mean_val,
            "p90": np.percentile(values, 90),
            "p99": np.percentile(values, 99),
            "cv": cv,
        }

    def calculate_mean_and_cv(self, values: List[float]) -> Dict[str, float]:
        """Calculate mean and CV (for token metrics)"""
        if not values:
            return {"mean": 0, "cv": 0}

        mean_val = np.mean(values)
        std_val = np.std(values, ddof=1) if len(values) > 1 else 0
        cv = (std_val / mean_val * 100) if mean_val > 0 else 0

        return {"mean": mean_val, "cv": cv}

    def generate_task_time_report(self, output_file: str):
        """
        Generate CSV report for task completion time grouped by task, architecture, and model
        """
        # Group data by task, architecture, model
        grouped = defaultdict(list)
        for parser in self.data:
            key = (parser.task, parser.architecture, parser.model)
            if parser.total_time:
                grouped[key].append(parser.total_time)

        # Calculate statistics
        results = []
        for (task, arch, model), times in grouped.items():
            stats = self.calculate_statistics(times)
            # Calculate throughput: tasks per hour
            throughput = 3600 / stats["mean"] if stats["mean"] > 0 else 0
            results.append(
                {
                    "task": task,
                    "architecture": arch,
                    "model": model,
                    "count": len(times),
                    "mean_time": stats["mean"],
                    "p90_time": stats["p90"],
                    "p99_time": stats["p99"],
                    "cv_time": stats["cv"],
                    "throughput_tasks_per_hour": throughput,
                }
            )

        # Sort by task, architecture, model
        results.sort(key=lambda x: (x["task"], x["architecture"], x["model"]))

        # Write to CSV
        with open(output_file, "w", newline="", encoding="utf-8") as f:
            writer = csv.DictWriter(
                f,
                fieldnames=[
                    "task",
                    "architecture",
                    "model",
                    "count",
                    "mean_time",
                    "p90_time",
                    "p99_time",
                    "cv_time",
                    "throughput_tasks_per_hour",
                ],
            )
            writer.writeheader()
            writer.writerows(results)

        logger.info(f"Task time report written to {output_file}")

    def generate_agent_time_report(self, output_file: str):
        """
        Generate CSV report for agent time grouped by task, architecture, agent, and model
        """
        # Group data by task, architecture, agent, model
        grouped = defaultdict(list)
        for parser in self.data:
            for agent_name, stats in parser.agent_stats.items():
                if stats["time"]:
                    key = (parser.task, parser.architecture, agent_name, parser.model)
                    grouped[key].extend(stats["time"])

        # Calculate statistics
        results = []
        for (task, arch, agent, model), times in grouped.items():
            stats = self.calculate_statistics(times)
            results.append(
                {
                    "task": task,
                    "architecture": arch,
                    "agent": agent,
                    "model": model,
                    "count": len(times),
                    "mean_time": stats["mean"],
                    "p90_time": stats["p90"],
                    "p99_time": stats["p99"],
                    "cv_time": stats["cv"],
                }
            )

        # Sort by task, architecture, agent, model
        results.sort(
            key=lambda x: (x["task"], x["architecture"], x["agent"], x["model"])
        )

        # Write to CSV
        with open(output_file, "w", newline="", encoding="utf-8") as f:
            writer = csv.DictWriter(
                f,
                fieldnames=[
                    "task",
                    "architecture",
                    "agent",
                    "model",
                    "count",
                    "mean_time",
                    "p90_time",
                    "p99_time",
                    "cv_time",
                ],
            )
            writer.writeheader()
            writer.writerows(results)

        logger.info(f"Agent time report written to {output_file}")

    def generate_task_token_report(self, output_file: str):
        """
        Generate CSV report for task token usage grouped by task, architecture, and model
        """
        # Group data by task, architecture, model
        grouped = defaultdict(
            lambda: {
                "input": [],
                "output": [],
                "reasoning": [],
                "result": [],
                "total": [],
            }
        )

        for parser in self.data:
            key = (parser.task, parser.architecture, parser.model)
            if parser.total_tokens["total"] > 0:
                for token_type in ["input", "output", "reasoning", "result", "total"]:
                    grouped[key][token_type].append(parser.total_tokens[token_type])

        # Calculate statistics
        results = []
        for (task, arch, model), token_data in grouped.items():
            if not token_data["total"]:
                continue

            result = {
                "task": task,
                "architecture": arch,
                "model": model,
                "count": len(token_data["total"]),
            }

            # Token statistics: mean and CV
            for token_type in ["input", "output", "reasoning", "result", "total"]:
                token_stats = self.calculate_mean_and_cv(token_data[token_type])
                result[f"mean_{token_type}"] = token_stats["mean"]
                result[f"cv_{token_type}"] = token_stats["cv"]

            results.append(result)

        # Sort by task, architecture, model
        results.sort(key=lambda x: (x["task"], x["architecture"], x["model"]))

        # Write to CSV
        fieldnames = ["task", "architecture", "model", "count"]
        for token_type in ["input", "output", "reasoning", "result", "total"]:
            fieldnames.extend([f"mean_{token_type}", f"cv_{token_type}"])

        with open(output_file, "w", newline="", encoding="utf-8") as f:
            writer = csv.DictWriter(f, fieldnames=fieldnames)
            writer.writeheader()
            writer.writerows(results)

        logger.info(f"Task token report written to {output_file}")

    def generate_agent_token_report(self, output_file: str):
        """
        Generate CSV report for agent token usage grouped by task, architecture, agent, and model
        """
        # Group data by task, architecture, agent, model
        grouped = defaultdict(
            lambda: {
                "input": [],
                "output": [],
                "reasoning": [],
                "result": [],
                "total": [],
            }
        )

        for parser in self.data:
            for agent_name, stats in parser.agent_stats.items():
                if stats["tokens"]["total"]:
                    key = (parser.task, parser.architecture, agent_name, parser.model)
                    for token_type in [
                        "input",
                        "output",
                        "reasoning",
                        "result",
                        "total",
                    ]:
                        grouped[key][token_type].extend(stats["tokens"][token_type])

        # Calculate statistics
        results = []
        for (task, arch, agent, model), token_data in grouped.items():
            if not token_data["total"]:
                continue

            result = {
                "task": task,
                "architecture": arch,
                "agent": agent,
                "model": model,
                "count": len(token_data["total"]),
            }

            # Token statistics: mean and CV
            for token_type in ["input", "output", "reasoning", "result", "total"]:
                token_stats = self.calculate_mean_and_cv(token_data[token_type])
                result[f"mean_{token_type}"] = token_stats["mean"]
                result[f"cv_{token_type}"] = token_stats["cv"]

            results.append(result)

        # Sort by task, architecture, agent, model
        results.sort(
            key=lambda x: (x["task"], x["architecture"], x["agent"], x["model"])
        )

        # Write to CSV
        fieldnames = ["task", "architecture", "agent", "model", "count"]
        for token_type in ["input", "output", "reasoning", "result", "total"]:
            fieldnames.extend([f"mean_{token_type}", f"cv_{token_type}"])

        with open(output_file, "w", newline="", encoding="utf-8") as f:
            writer = csv.DictWriter(f, fieldnames=fieldnames)
            writer.writeheader()
            writer.writerows(results)

        logger.info(f"Agent token report written to {output_file}")

    def check_has_retry(self, parser: ExecutionPathParser) -> bool:
        """
        Check if execution_path.md contains retry markers like [RETRY1], [BUSINESS-RETRY] etc.
        Returns True if retry markers found, False otherwise
        """
        try:
            with open(parser.file_path, "r", encoding="utf-8") as f:
                content = f.read()
                # Pattern to match retry markers: [RETRYX], [BUSINESS-RETRY], etc.
                # Matches: [RETRY<number>] or [<any_text>RETRY<any_text>]
                retry_patterns = [
                    r"\[RETRY\d+\]",  # [RETRY1], [RETRY2], etc.
                    r"\[.*?RETRY.*?\]",  # [BUSINESS-RETRY], [XXXRETRYXXX], etc.
                ]

                for pattern in retry_patterns:
                    if re.search(pattern, content):
                        return True
                return False
        except Exception as e:
            logger.error(f"Error checking retry markers for {parser.file_path}: {e}")
            return False

    def check_task_success(self, parser: ExecutionPathParser) -> str:
        """
        Check if a task execution was successful based on task-specific criteria
        Returns 'success' or 'fail'
        """
        try:
            # Get the directory containing execution_path.md
            session_dir = Path(parser.file_path).parent
            task = parser.task

            if task.endswith("-H_A2A"):
                task = task[: -len("-H_A2A")]

            task_aliases = {
                "BookWriter": "write_a_book_with_flows",
                "SQLAssistant": "SQL_assistant",
                "SocialMediaManager": "self_evaluation_loop_flow",
                "LandingPageGenerator": "landing_page_generator",
                "RecruitmentAssistant": "intelligent_recruitment_platform",
                "EmailResponder": "EmailResponder",
                "GameBuilder": "GameBuilder",
                "MarkdownValidator": "MarkdownValidator",
            }
            task = task_aliases.get(task, task)

            # 1. write_a_book_with_flows
            if task == "write_a_book_with_flows":
                chapters_dir = session_dir / "chapters"
                if chapters_dir.exists() and chapters_dir.is_dir():
                    md_files = list(chapters_dir.glob("*.md"))
                    if len(md_files) >= 1:
                        return "success"
                return "fail"

            # 2. SQL_assistant
            elif task == "SQL_assistant":
                with open(parser.file_path, "r", encoding="utf-8") as f:
                    content = f.read()
                    if re.search(r"\[Tool\]\s+get_database_schema", content):
                        return "success"
                return "fail"

            # 3. self_evaluation_loop_flow
            elif task == "self_evaluation_loop_flow":
                metadata_file = session_dir / "metadata.json"
                if metadata_file.exists():
                    with open(metadata_file, "r", encoding="utf-8") as f:
                        metadata = json.load(f)
                        if metadata.get("status") == "success":
                            return "success"
                return "fail"

            # 4. MarkdownValidator
            elif task == "MarkdownValidator":
                execution_info_file = session_dir / "execution_info.json"
                if execution_info_file.exists():
                    with open(execution_info_file, "r", encoding="utf-8") as f:
                        exec_info = json.load(f)
                        if exec_info.get("success", False):
                            return "success"
                return "fail"

            # 5. landing_page_generator
            elif task == "landing_page_generator":
                html_validation_file = session_dir / "html_validation.json"
                if html_validation_file.exists():
                    with open(html_validation_file, "r", encoding="utf-8") as f:
                        validation = json.load(f)
                        if validation.get("file_exists", False):
                            return "success"
                return "fail"

            # 6. intelligent_recruitment_platform
            elif task == "intelligent_recruitment_platform":
                reports_dir = session_dir / "reports"
                if reports_dir.exists() and reports_dir.is_dir():
                    md_files = list(reports_dir.glob("*.md"))
                    if len(md_files) >= 2:
                        return "success"
                return "fail"

            # 7. GameBuilder
            elif task == "GameBuilder":
                validation_result_file = session_dir / "validation_result.json"
                if validation_result_file.exists():
                    with open(validation_result_file, "r", encoding="utf-8") as f:
                        validation = json.load(f)
                        if validation.get("validation_successful", False):
                            return "success"
                return "fail"

            # 8. EmailResponder
            elif task == "EmailResponder":
                execution_log_file = session_dir / "execution_log.json"
                if execution_log_file.exists():
                    with open(execution_log_file, "r", encoding="utf-8") as f:
                        exec_log = json.load(f)
                        if exec_log.get("success", False):
                            return "success"
                return "fail"

            # Unknown task
            else:
                logger.warning(f"Unknown task type for success check: {task}")
                return "fail"

        except Exception as e:
            logger.error(f"Error checking task success for {parser.file_path}: {e}")
            return "error"

    def generate_task_token_details(self, output_file: str):
        """
        Generate detailed CSV report with token usage for each individual execution_path.md file
        """
        results = []

        for parser in self.data:
            if parser.total_tokens["total"] > 0:
                # Check task success status
                status = self.check_task_success(parser)

                # Check if execution contains retry markers
                has_retry = self.check_has_retry(parser)

                result = {
                    "file_path": parser.file_path,
                    "task": parser.task,
                    "architecture": parser.architecture,
                    "model": parser.model,
                    "status": status,
                    "with_retry": str(
                        has_retry
                    ).lower(),  # Convert to 'true' or 'false'
                    "input_tokens": parser.total_tokens["input"],
                    "output_tokens": parser.total_tokens["output"],
                    "reasoning_tokens": parser.total_tokens["reasoning"],
                    "result_tokens": parser.total_tokens["result"],
                    "total_tokens": parser.total_tokens["total"],
                }
                results.append(result)

        # Sort by task, architecture, model, and file path
        results.sort(
            key=lambda x: (x["task"], x["architecture"], x["model"], x["file_path"])
        )

        # Write to CSV
        with open(output_file, "w", newline="", encoding="utf-8") as f:
            writer = csv.DictWriter(
                f,
                fieldnames=[
                    "file_path",
                    "task",
                    "architecture",
                    "model",
                    "status",
                    "with_retry",
                    "input_tokens",
                    "output_tokens",
                    "reasoning_tokens",
                    "result_tokens",
                    "total_tokens",
                ],
            )
            writer.writeheader()
            writer.writerows(results)

        logger.info(f"Detailed task token report written to {output_file}")


def main():
    """Main execution function"""
    # Set paths
    results_dir = "/Users/wzr/TOSEM-2025/RESULTS"
    output_dir = "/Users/wzr/TOSEM-2025/RESULTS/RQ3/performance_reports"

    # Create output directory if it doesn't exist
    os.makedirs(output_dir, exist_ok=True)

    # Initialize analyzer
    logger.info("Starting performance analysis...")
    analyzer = PerformanceAnalyzer(results_dir)

    # Parse all files
    logger.info("Parsing execution path files...")
    analyzer.parse_all_files()

    # Generate reports
    logger.info("Generating reports...")

    # Time reports
    analyzer.generate_task_time_report(
        os.path.join(output_dir, "task_time_statistics.csv")
    )
    analyzer.generate_agent_time_report(
        os.path.join(output_dir, "agent_time_statistics.csv")
    )

    # Token reports
    analyzer.generate_task_token_report(
        os.path.join(output_dir, "task_token_statistics.csv")
    )
    analyzer.generate_agent_token_report(
        os.path.join(output_dir, "agent_token_statistics.csv")
    )

    # Detailed token report (per file)
    analyzer.generate_task_token_details(
        os.path.join(output_dir, "task_token_statistics-DETAILS.csv")
    )

    logger.info("Analysis complete!")
    logger.info(f"Total files processed: {len(analyzer.data)}")
    logger.info(f"Failed files: {len(analyzer.failed_files)}")
    logger.info(f"Reports saved to: {output_dir}")


if __name__ == "__main__":
    main()