Run Output

The Statistics table gives one row per sample label plus a Total row: Label, #Samples, FAIL, Error %, Average, Min, Max, Median, three percentile columns, Transactions/s, and Received and Sent KB/sec.

JMeter's default CSV JTL carries seventeen columns: timeStamp, elapsed, label, responseCode, responseMessage, threadName, dataType, success, failureMessage, bytes, sentBytes, grpThreads, allThreads, URL, Latency, IdleTime and Connect.

The Backend Listener. Pick a BackendListenerClient implementation - InfluxdbBackendListenerClient or GraphiteBackendListenerClient - then fill its Parameters table with the target URL or host and port, and metrics leave the injector during the run.

Run the report generator on its own with -g: jmeter -g last-night.jtl -o dashboard. The flag takes an existing results file, so no plan is loaded and no load is generated.

All three are millisecond timings taken from the same instant the sample started. Connect stops when the connection is established, Latency when the first response arrives, elapsed when the last response byte arrives.

Setting summaryOnly to true switches per-sampler metrics off entirely, so samplersList is never consulted and only the cumulated series is sent. The filter has no effect until summaryOnly is false.

The two surfaces use different percentile estimators. JMeter's HTML dashboard uses the commons-math3 LEGACY type, which interpolates between neighbouring samples; the Aggregate Report listener returns an actually observed elapsed time. Set backend_metrics_percentile_estimator=R_3 to align them.

The three properties aggregate_rpt_pct1, aggregate_rpt_pct2 and aggregate_rpt_pct3, defaulting to 90, 95 and 99. Each takes a float between 0 and 100, so 99.9 is valid, and they move the GUI listeners and the HTML dashboard together.

The folder given to -o already contains files. JMeter checks the report destination before the run starts and refuses rather than mix a new dashboard into an old one. The -f flag clears it first.

T holds jmeter.reportgenerator.apdex_satisfied_threshold, 500 ms by default; F holds apdex_tolerated_threshold, 1500 ms by default. Despite the headings, T is the satisfied bound and F the tolerated bound, both in milliseconds.

Average and the 99th pct column are computed from exactly the same list of elapsed times: every sample carrying that label inside the report's range, failures included, over the whole run with no time dimension.

The jmeter.save.saveservice.output_format property decides it, and it defaults to csv. Only the exact lowercase value xml selects XML; anything else logs a warning and falls back to CSV. The file name is irrelevant.

Sub-samples get rows of their own. jmeter.save.saveservice.subresults defaults to true, so every child result a sampler produced is written as an extra flat row after its parent, relabelled parentLabel-0, parentLabel-1 and so on.

A Summariser element, known in the GUI as Generate Summary Results. JMeter auto-creates one in CLI mode named by the summariser.name property, shipped as summary; + lines are the delta since the last report and = lines the running total.

Lower jmeter.reportgenerator.overall_granularity. It defaults to 60000 milliseconds, so every over-time graph gets one point per minute. Set it in user.properties, keep it above 1000, and regenerate the existing results file.

JMeter divides each row's sample count by that label's own span, from its first sample's start to its last sample's end. Rows therefore use different, overlapping denominators, so their rates are not additive.

Both symptoms come from save properties on the injector that wrote it: print_field_names was false, and timestamp_format was a date pattern instead of the default ms. Rebuild the column list from JMeter's fixed field order, then confirm the writing machine's time zone.

Over a fixed-size sliding window of the most recent response times - 100 values by default, set by backend_metrics_window - and never over the interval just elapsed. That window is not cleared between flushes.

The dashboard's percentile columns run over a sliding window of the most recent samples, sized by jmeter.reportgenerator.statistic_window (default 20,000 per row). Min, Max and Average are not windowed, so they still see the whole run.

Start from summaryOnly and the sampler filter: they decide whether one series or one per label leaves the injector. Then settle the percentile list, the flush interval and a label naming rule, and hold all of them in properties rather than in each plan.

Ship one properties file with every JMeter install carrying the estimator, the report window and the three percentile levels. All three are read at startup, and the JVM generating the report needs them too.

The key is backend_metrics_percentile_estimator, with no jmeter.reportgenerator prefix. It accepts a commons-math3 Percentile.EstimationType constant name: LEGACY (the default) or R_1 through R_9. Nothing else parses.

Set jmeter.reportgenerator.exporter.html.series_filter to a regular expression matching the series you want, ending it with (-success|-failure)?$ so the per-second graphs still match. An empty value, the default, filters nothing.

The Errors table lists Type of error, Number of errors, % in errors and % in all samples, one row per distinct failure key — unlike the Statistics and APDEX tables it has no Total row. The key is normally the response code and message.

Connect is only timed when a connection is actually established, and only by the sampler families that record it. A reused keep-alive connection does no connecting, so the column is a truthful zero rather than a missing value.

The sampling thread that produced the result blocks until the worker frees a slot. Nothing is dropped; the delay lands on the test thread, and JMeter logs a queue-wait warning once the test ends.