What is coordinated omission?
Section titled “What is coordinated omission?”Coordinated Omission is a measurement bias that occurs when a load generator schedules the next request at a fixed interval (e.g., every 20 ms for 50 RPS) but fails to account for the time already spent waiting on the previous request.
When the server degrades:
- Threads get “stuck” waiting for slow responses.
- Fewer requests are sent than the target rate.
- The “missing” requests are silently dropped - never queued.
- Measured throughput drops, but response times look artificially low.
Example: You target 50 RPS with 100 threads. Server degrades to 5-second response time. Instead of 50 RPS, you get ~20 RPS (100 threads / 5s). Your throughput drops by 60%, and your average response time doesn’t reflect what real users would experience.
How the correction works
Section titled “How the correction works”The calculator uses the queue-depth model from Gil Tene’s talk How NOT to Measure Latency:
correction_factor = target_RPS / actual_RPSlost_requests = (target_RPS - actual_RPS) × duration_secondsqueue_depth = (target_RPS - actual_RPS) × (avg_RT_ms / 1000)corrected_RT = avg_RT_ms + (queue_depth × avg_RT_ms)needed_threads = target_RPS × (avg_RT_ms / 1000)When target_RPS > actual_RPS, the correction factor tells you how much your
measured metrics are skewed. A factor of 2.5x means your real response time is
roughly 2.5× what JMeter reported.
Example
Section titled “Example”| Target RPS | Actual RPS | Avg RT | Duration | Correction | Lost reqs | Corrected RT |
|---|---|---|---|---|---|---|
| 50 | 20 | 5,000ms | 60s | 2.5x | 1,800 | ~12,500ms |
| 100 | 80 | 3,000ms | 120s | 1.25x | 2,400 | ~3,750ms |
| 200 | 200 | 500ms | 60s | 1.0x | 0 | 500ms |
What this does not include
Section titled “What this does not include”- Percentile correction - the queue-depth model corrects averages; for percentile correction, use a proper histogram (HdrHistogram) in your test.
- Think time / pacing - real users pause between actions; this calculator assumes continuous load.
- Server-side queuing - if the server itself queues requests, the corrected RT is a lower bound on what users experience.
- Distributed test aggregation - run this per engine and combine results.
Recommended workflow
Section titled “Recommended workflow”- Run your test and record target RPS, actual RPS, avg response time, and duration.
- Use this calculator to compute the correction factor and lost requests.
- If the factor is > 1.25, your results are significantly skewed.
- Re-run with more threads (use the Thread Calculator) or lower the target RPS.
- For accurate percentiles, instrument with HdrHistogram in your JMeter test.