Package org.acemq.workloads.metrics

Class LatencyRecorder

java.lang.Object
org.acemq.workloads.metrics.LatencyRecorder
public final class LatencyRecorder extends Object
Records latencies, from many threads, without distorting them.

Why a histogram and not a list

Two reasons, and both are fatal to the alternative.

Percentiles cannot be averaged. The mean of each thread's p99 is not the p99, and neither is the mean of each second's p99. Any design that summarises per thread or per interval and combines the summaries afterwards produces a number that is not a percentile of anything. The samples have to meet in one place.

A list does not fit. At 300,000 messages a second for five minutes that is ninety million longs — 720MB, allocated during the measurement, which the garbage collector then pays for in exactly the pauses being measured.

ConcurrentHistogram solves both: fixed memory, lock-free recording, and exact percentiles to the configured precision.

  • Constructor Details

    • LatencyRecorder

      public LatencyRecorder(String name)
      Parameters:
      name - what is being measured, for the report

  • Method Details

    • record

      public void record(long nanos)
      Parameters:
      nanos - one observation

    • summary

      public LatencySummary summary()
      Returns:
      a snapshot of everything recorded so far

    • count

      public long count()
      Returns:
      how many observations have been recorded

    • reset

      public void reset()
      Discards everything recorded so far, which is how warm-up is excluded.

    • toString

      public String toString()
      Overrides:
      toString in class Object