분산 결과 집계 및 동기화

In distributed load testing many generators run in parallel. Each produces its own slice of results. To understand total system behavior you must aggregate these slices into a single, consistent view.

A single node might report 500 requests/sec, but with eight nodes the real throughput is roughly 4000 requests/sec. Looking at one node alone underestimates load and can hide saturation of the target system.

If generators have drifting clocks, merged time series will be misaligned and percentiles meaningless. Always run NTP so all nodes share a common time base before testing.

The cleanest way to aggregate is to stream every node's metrics to one backend. Both JMeter and k6 can push to time-series databases such as InfluxDB, where data is merged automatically by timestamp and tags.

Run each k6 instance with an output flag pointing to the shared backend. Tag each run with its node so you can still drill down per generator.

JMeter writes per-node JTL result files. You can combine them by concatenating (keeping one header) and then loading the merged file into the JMeter GUI or a report generator.

You cannot average per-node percentiles to get a global percentile. Correct aggregation requires the raw response times from all nodes combined, then computing the percentile over the full dataset.

Once results are merged, JMeter can produce an HTML dashboard from the combined JTL, giving one report for the whole distributed run.

Trim the warm-up and ramp-down so all nodes contribute only their steady-state window. Comparing overlapping time ranges keeps throughput and latency numbers honest.

With data in InfluxDB, a Grafana dashboard can sum throughput across nodes and chart global percentiles in real time, giving one live picture of the distributed test.

Distributed runs must start together. Use an orchestrator or a shared trigger so every generator begins its ramp at the same instant, otherwise their time windows never overlap cleanly.

Check your understanding of distributed aggregation.

You learned to make distributed results coherent.

• Synchronize clocks with NTP before testing.
• Stream to a central backend or merge JTL files carefully.
• Recompute percentiles over combined raw data, never by averaging node percentiles.