Performance

Envoy is architected to optimize scalability and resource utilization by running an event loop on a small number of threads. The “main” thread is responsible for control plane processing, and each “worker” thread handles a portion of the data plane processing. Envoy exposes two statistics to monitor performance of the event loops on all these threads.

  • Loop duration: Some amount of processing is done on each iteration of the event loop. This amount will naturally vary with changes in load. However, if one or more threads have an unusually long-tailed loop duration, it may indicate a performance issue. For example, work might not be distributed fairly across the worker threads, or there may be a long blocking operation in an extension that’s impeding progress.
  • Poll delay: On each iteration of the event loop, the event dispatcher polls for I/O events and “wakes up” either when some I/O events are ready to be processed or when a timeout fires, whichever occurs first. In the case of a timeout, we can measure the difference between the expected wakeup time and the actual wakeup time after polling; this difference is called the “poll delay.” It’s normal to see some small poll delay, usually equal to the kernel scheduler’s “time slice” or “quantum”—this depends on the specific operating system on which Envoy is running—but if this number elevates substantially above its normal observed baseline, it likely indicates kernel scheduler delays.

These statistics can be enabled by setting enable_dispatcher_stats to true.

Warning

Note that enabling dispatcher stats records a value for each iteration of the event loop on every thread. This should normally be minimal overhead, but when using statsd, it will send each observed value over the wire individually because the statsd protocol doesn’t have any way to represent a histogram summary. Be aware that this can be a very large volume of data.

Statistics

The event dispatcher for the main thread has a statistics tree rooted at server.dispatcher., and the event dispatcher for each worker thread has a statistics tree rooted at listener_manager.worker_.dispatcher., each with the following statistics:

NameTypeDescription
loop_duration_usHistogramEvent loop durations in microseconds
poll_delay_usHistogramPolling delays in microseconds

Note that any auxiliary threads are not included here.