• Understand the issue: The clock starts before state() and setup() methods complete, causing setup time to be included in throughput calculations
• Implement fix: Use a barrier to synchronize all workers, start the clock after all setup() and warmup methods complete
• Modify Clock to support paused initialization with new_paused()
• Update Runner to use barrier synchronization and start clock after setup and warmup
• Update cli.rs to create clock in paused state initially
• Test the changes with cargo test and manual verification
• Run code review and address feedback - changed Clock::pause() and Clock::resume() to take &self instead of &mut self
• Refactor: Extract common Clock construction logic into private new() method to eliminate duplicate code between start_at() and new_paused()

Summary

This PR fixes issue #20 where the load test summary includes setup time, leading to inaccurate throughput calculations.

Changes Made

1. Added Clock::new_paused() to create a clock in paused state initially
2. Added a barrier synchronization using tokio::sync::Barrier to wait for all workers to complete setup and warmup
3. The clock is only resumed when all workers are ready to start the main benchmark iterations
4. Changed Clock::pause() and Clock::resume() to take &self (more idiomatic since state is mutex-protected)
5. Refactored Clock::start_at() and Clock::new_paused() to share common construction logic via private new() method

Testing

• All existing tests pass
• Manual verification with warmup example shows correct timing:
  • With warmup: 0.08s for main iterations only
  • Without warmup: 1.54s (includes slow "cold start" iterations)
• Duration parameter works correctly (runs for exactly the specified duration)

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.